cute: remove redundant subprocess-autotune DSL recompiles

[calebmkim]

Stacked PRs:

• [autotuner][cute] add formula-based tcgen05 matmul seed heuristic #3152
• [autotuner][cute] rename ab_stages "three" symbols that are no longer 3-only #3151
• [autotuner][cute] widen the tcgen05 matmul search space: block_n=128 cluster_m=2 + ab=4 #3150
• [autotuner][cute] repair partial FFI candidates away from the seed, not onto it #3167
• [autotuner][cute] demote tcgen05_tvm_ffi_launch to a seed-only knob #3149
• ->cute: remove redundant subprocess-autotune DSL recompiles #3148

---

cute: remove redundant subprocess-autotune DSL recompiles

The killable-worker autotune path cold-compiled the CuTe DSL kernel ~twice per
config (separate bench + accuracy jobs) and the worker disk cache was disabled
(cache_key=None), so a fixed budget searched far fewer distinct configs under
subprocess than in-process. A cold cutlass-DSL compile is ~3.5s vs a ~0.05s warm
disk reload, so the wasted recompiles dominated.

Fixes: (1) BenchmarkAndAccuracyJob compiles the source once for both bench and
accuracy (behind HELION_FUSED_ACCURACY_CHECK, default off); (2)
_reattach_cute_source_hash restores the worker disk-cache key so repeats are warm
reloads.

Effect (single B200, effort=full, seed=2000, subprocess=ON, fp8 4096^3):
distinct configs in a 300 s budget: 44 -> 83 (+89%; 1 -> 2 generations). Inert on
Triton and in-process. (The 44 "before" arm is the parent commit d51ece6d8^, which
predates both fixes; "after" is this commit + HELION_FUSED_ACCURACY_CHECK=1.)

Compile-COST only: does not change the benchmark timer or selection. Verified no
selection regression from this change: on a >=2-generation autotune run, the final
winner ([256,256,128] cm2 ab6) was independently cold-L2 cudagraph-timed against the
heuristic seed and was within 30% of it (winner/seed = 1.00x, PASS -- the winner is
the seed config) -- i.e. the fix does not cause the tuner to crown a config markedly
worse than a known-good one.

Gating: the two fixes have DIFFERENT blast radii. (1) the fused job is behind
HELION_FUSED_ACCURACY_CHECK, default off. (2) reattach_cute_source_hash is
UNCONDITIONAL and takes effect on every worker load, i.e. on the default-ON
subprocess-benchmark path. That is deliberate: the parent already stamps this hash
and already uses the on-disk DSL cache, so the worker's cache_key=None was a defect,
not a policy -- and a wrong stamp can only MISS the cache (cold recompile), never
produce a false hit, because the key also carries the full input specialization,
launch shape, compile options, CUTE_DSL* env, and cutlass version. Its correctness
does rest on the timer absorbing the first call: both do_bench (benchmarking.py, fn()

• sync before capture) and do_bench_generic (fn() + sync before the estimate loop)
  already do, so the compile-or-reload cost never enters the timed samples. The NOTE
  added to do_bench_generic marks that ordering as load-bearing for future editors --
  if warmup ever moved after the timed loop, the reattach would turn a uniform
  first-call penalty into a non-uniform one (cold for first-seen configs, warm for
  repeats), which would bias selection where a uniform penalty largely cancels.

Timeout accounting: the fused job runs bench AND the accuracy check under a SINGLE
autotune_benchmark_timeout window, where the legacy two-job path gave each step its
own. The extra work inside that window is one kernel launch plus an assert_close --
not a second compile, which is the point of the change -- so the narrowing is small
next to a ~3.5 s cold compile, but it is not zero: a config near the timeout boundary
has slightly less headroom than before. Not compensated for deliberately (doubling
the window would also double what a genuinely hung config gets to hang for, weakening
the killable-worker hang protection).

Adaptation: zero-arg synchronize_device() (PR #3026). Refactor: 3 copies of
subprocess job-error handling -> _handle_subprocess_job_error.

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com