perf: IDOT_TEAM=1 fuses three OpenMP regions into one team per expert…

[KyleSanderson]

Summary

perf: IDOT_TEAM=1 fuses three OpenMP regions into one team per expert forward
Eliminates 2x team barrier spin (gomp_team_barrier_wait_end + gomp_barrier_wait_end)
per expert forward. VTune measured 168.3 CPU-s of OpenMP imbalance/serial spin.
A/B test on GLM-5.2 real workload (--ram 8):
prefill: 73.17s -> 71.20s (-2.7%)
decode: 75.20s -> 62.07s (-17.5%)
No regression in expert-matmul IPC or output correctness.
Perf c2c confirmed no material false sharing (max HITM 2.15%).

Validation

• make -C c check
• CUDA changes were tested with make -C c cuda-test (if applicable)
• Performance claims include hardware, commands, and repeatable measurements

Compatibility

• The default CPU build remains dependency-free
• No model files, generated binaries, or benchmark artifacts are included

[JustVugg]

Status: every Linux and ARM job is red on this branch (13 of 26 checks), including the token-exact oracles, so it cannot be reviewed for the speed claim yet. The change itself is opt-in and looks numerically neutral (same per-element arithmetic, collapse(2) over the same loops), so the failures are probably the build or a stale base: please rebase onto dev and look at one red log. With CI green, the number I would need is an A/B on one binary (IDOT_TEAM=0 vs 1, same prompt, tok/s), since the getenv sits on a per-call path.