perf(sm90): size grouped W4A8 tiles per expert

[LopezCastroRoberto]

In #124, grouped-contiguous W4A8 still takes its M tile from the generic seed heuristic, which sizes tiles from total rows as for a dense GEMM. This PR sizes the tile per expert instead, in two commits.

1. Row-major tile model (17d138d). In #124 the packed N16 schedule is applied only where the seed tile reaches M128. This causes two problems:

• Seed-schedule holes. For total rows in (4416, 6144], (8512, 9472], (12160, 12416] and (16000, 16192] (≈ M 552-768, 1064-1184, 1520-1552 and 2000-2024 tokens at top-k 8), the seed tile is M96-M120, so these sizes run the generic warp-N32 schedule. That is 1.3-1.5× slower than the N16 schedule (M=768, 1152 and 1536 below).
• M128 splits unevenly routed experts. With random routing, experts that land slightly over 128 or 256 rows need an extra, mostly empty tile. At M=512 that is 45 tiles instead of 32 (+41%), and at M=1024 it is 81 instead of 64 (+27%).

On H200, row-major now uses the same expected-work tile model as the M-major path (_w4a8_tile_m_for_expert_rows): the model tile when it is ≤ M160, M176 only while one tile covers the expert (≤ 176 rows), M128 beyond that, and no raster_group_m.

2. Small-expert seed tile, both layouts (3be4376). At ≤ 96 rows per expert, both scale layouts keep the seed schedule, but its M tile now comes from rows per expert (rows + √rows, rounded up to 8), with a (256, 128) block N/K when N is a multiple of 256. Before this change:

• The row-major seed tile was about twice the rows of one expert; on down-proj, the (32, 256, 256) tile for 320-512 rows was 1.35× slower than the tiles on either side.
• M-major used [Perf] Optimize Hopper MXFP4 × FP8_BLOCK grouped MoE prefill #95's packed N16 schedule with an M64+ tile, which was 1.03-1.55× slower than the per-expert seed tile at these sizes.

The 96 cutoff comes from a dense M=8-512 sweep (step 8): the crossover with the N16 tile model sits at 96-100 rows per expert for both layouts, both shapes and both routings.

Other SM90 devices keep the existing selection, as for the M-major tile model. No CUDA changes. Accuracy against the KernelTestRunner FP32 reference is unchanged (both layouts, M=8-512). test_grouped_w4a8_ranges_match_direct_selection is updated, because small-expert M-major ranges on H200 now use the seed schedule.

Results (H200, median of 3; M-major includes #126's split cutoff):

python benchmarks/bench_humming.py --shape_n 4096 --shape_k 6144 \
  --a_dtype float8e4m3 --b_dtype float4e2m1 --c_dtype bfloat16 --bs_dtype float8e8m0 \
  --input_scale_group_size 128 --weight_scale_group_size 32 \
  --gemm_type grouped_contiguous --num_experts 32 --top_k 8 [--use_m_major_input_scale]

Gate/up (N=4096, K=6144), default (unbalanced) routing

M	row-major #124	row-major #124 + #126 + #129	speedup	M-major #124 + #126 + #129	row-major vs M-major
64	0.1722	0.1540	1.12×	0.1537	1.00×
256	0.2312	0.2180	1.06×	0.2214	1.02×
512	0.3888	0.3104	1.25×	0.3163	1.02×
640	0.4739	0.4212	1.13×	0.4044	0.96×
768	0.6615	0.5146	1.29×	0.5642	1.10×
1024	0.6812	0.5977	1.14×	0.5961	1.00×
1152	0.9765	0.6524	1.50×	0.6464	0.99×
1536	1.2784	0.8778	1.46×	0.8798	1.00×
2048	1.1710	1.1709	1.00×	1.1192	0.96×
4096	2.1756	2.1796	1.00×	2.1108	0.97×
16384	8.2034	8.2021	1.00×	7.9913	0.97×

Down (N=6144, K=2048), default (unbalanced) routing

M	row-major #124	row-major #124 + #126 + #129	speedup	M-major #124 + #126 + #129	row-major vs M-major
64	0.1277	0.0938	1.36×	0.0938	1.00×
256	0.1267	0.1277	0.99×	0.1308	1.02×
512	0.2197	0.1764	1.25×	0.1777	1.01×
640	0.2678	0.2276	1.18×	0.2218	0.97×
768	0.3621	0.2811	1.29×	0.2808	1.00×
1024	0.3803	0.3326	1.14×	0.3369	1.01×
1152	0.5323	0.3693	1.44×	0.3627	0.98×
1536	0.6902	0.4824	1.43×	0.4951	1.03×
2048	0.6505	0.6529	1.00×	0.5938	0.91×
4096	1.2156	1.2186	1.00×	1.1721	0.96×
16384	4.5628	4.5630	1.00×	4.4139	0.97×

• M-major vs perf(sm90): apply the packed MXFP4 x FP8_BLOCK schedule to every GEMM type #124 + fix(tune): restore late AS and split issue/wait for M-major W4A8 grouped GEMM #126, from the small-expert seed tile at M ≤ 256: gate/up 1.18-1.55×, down 1.03-1.34×. M-major ≥ 512 is unchanged.
• Gain from the 96 cutoff over a 64-row cutoff: 1.02-1.28× at M=264-376 for both layouts (not shown in the tables).
• Layout parity: with this PR, row-major and M-major are within about ±3% of each other from M=8 to 512.
• Down M=2048: varies between ~0.553 and ~0.594 ms run to run in both layouts with identical configs, so treat that row as noise.
• Balanced routing, M=512/1024: the model's M144 tile is 5-7% slower than perf(sm90): apply the packed MXFP4 x FP8_BLOCK schedule to every GEMM type #124's M128 there, because M128 tiles come out exactly full. M-major makes the same choice.

Further experiments at lower Ms

Small M (seed path, ≤ 96 rows per expert), Gate/up (N=4096, K=6144), default (unbalanced) routing

M	row-major #124	row-major #124 + #126 + #129	speedup	M-major #124 + #126	M-major #124 + #126 + #129	speedup	row-major vs M-major (#129)
8	0.1445	0.1329	1.09×	0.2091	0.1348	1.55×	1.01×
16	0.1584	0.1412	1.12×	0.2104	0.1431	1.47×	1.01×
32	0.1669	0.1497	1.11×	0.2165	0.1494	1.45×	1.00×
48	0.1719	0.1492	1.15×	0.2185	0.1489	1.47×	1.00×
64	0.1722	0.1543	1.12×	0.2202	0.1535	1.43×	0.99×
96	0.1862	0.1574	1.18×	0.2199	0.1592	1.38×	1.01×
128	0.2010	0.1649	1.22×	0.2231	0.1663	1.34×	1.01×
192	0.1950	0.1877	1.04×	0.2250	0.1908	1.18×	1.02×
256	0.2312	0.2184	1.06×	0.2619	0.2214	1.18×	1.01×
264	0.2308	0.2107	1.10×	0.2606	0.2056	1.27×	0.98×
288	0.2471	0.2243	1.10×	0.2613	0.2166	1.21×	0.97×
320	0.2854	0.2336	1.22×	0.2654	0.2377	1.12×	1.02×
352	0.3400	0.2511	1.35×	0.2931	0.2545	1.15×	1.01×
384	0.2855	0.2883	0.99×	0.2943	0.2853	1.03×	0.99×

Small M (seed path, ≤ 96 rows per expert), Down (N=6144, K=2048), default (unbalanced) routing

M	row-major #124	row-major #124 + #126 + #129	speedup	M-major #124 + #126	M-major #124 + #126 + #129	speedup	row-major vs M-major (#129)
8	0.0791	0.0820	0.96×	0.1072	0.0835	1.28×	1.02×
16	0.0879	0.0859	1.02×	0.1138	0.0875	1.30×	1.02×
32	0.0944	0.0905	1.04×	0.1145	0.0900	1.27×	0.99×
48	0.1275	0.0909	1.40×	0.1154	0.0904	1.28×	0.99×
64	0.1277	0.0938	1.36×	0.1162	0.0942	1.23×	1.00×
96	0.1002	0.0962	1.04×	0.1176	0.0968	1.21×	1.01×
128	0.1075	0.0988	1.09×	0.1187	0.1001	1.19×	1.01×
192	0.1118	0.1128	0.99×	0.1205	0.1139	1.06×	1.01×
256	0.1267	0.1279	0.99×	0.1353	0.1310	1.03×	1.02×
264	0.1257	0.1167	1.08×	0.1344	0.1164	1.15×	1.00×
288	0.1376	0.1237	1.11×	0.1354	0.1215	1.11×	0.98×
320	0.1562	0.1302	1.20×	0.1370	0.1291	1.06×	0.99×
352	0.1884	0.1400	1.35×	0.1598	0.1431	1.12×	1.02×
384	0.1648	0.1591	1.04×	0.1610	0.1594	1.01×	1.00×