[Bench][Linear-Attn] Add Qwen GDN prefill benchmark surface

[superAngGao]

Stacked on tile-ai#1596. Addresses tile-ai#1597.

Summary

• Switch the Gated DeltaNet prefill benchmark surface to the FLA/Qwen BTHD contract.
• Add Qwen3.5-aligned long-context rows for S=32K/64K/128K and H=16/32/48/64 across fp16/bf16.
• Compare TileOps against FLA with output_final_state=True and surface speedup in profile_run.log.
• Keep nightly JUnit properties compatible while allowing profile_run.log to show extra metrics and GPU clocks.

Validation

• python -m py_compile benchmarks/ops/bench_gated_deltanet_prefill.py workloads/gated_deltanet.py benchmarks/benchmark_base.py tileops/perf/formulas.py
• python scripts/validate_manifest.py
• python -m pytest --collect-only benchmarks/ops/bench_gated_deltanet_prefill.py -q
• TMPDIR=/home/ga/tmp python -m pytest tests/ops/test_gated_deltanet_prefill.py -m smoke -q

Notes

• The local environment does not have FLA installed, so full benchmark execution was only checked for collection/skip behavior here.