[autotuner][cute] add formula-based tcgen05 matmul seed heuristic

[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

---

[autotuner][cute] add formula-based tcgen05 matmul seed heuristic

Replaces the special-case cluster_m=2 seed producers with one shape-aware analytical
formula (cute analog of the Triton B200 formula heuristic #3007):
f(M,K,N,dtype,epilogue,num_sm,smem_budget) -> Config, promoted to the no-autotune
default. Covers configs the old producers structurally could not emit: cluster_m=1
decode (M<=128, previously NO seed) and deep-AB compute (ab=6 vs the old ab=3).

Depends on PRs 1-3 below it (they widen the search surface / unblock injection); this
PR is the config PRODUCER on top of that surface.

STATUS: this is a STARTING POINT, not a finished heuristic. It is a first, deliberately
simple analytical formula (~5 knobs) that already covers regimes the old producers
could not and beats them on quality (below), but it is expected to be iterated on:
the classifier thresholds (e.g. the wave-fill / _MANY_WAVE crossovers), the l2_groupings
rule, and coverage of more epilogues/shapes are all approximations with known gaps (see
the F1 limitation below). Intent is to land a maintainable baseline that replaces the
special-case producers and then improve it incrementally, not to ship a final tuning oracle.

What this PR is responsible for (measured this run, single B200, cold-L2 cudagraph):

[A] Config quality -- the config the formula EMITS vs the old producers / an external
reference, benched directly (no autotuner):

• geomean(seed / old-producer) = 1.091 over the 23 curriculum shapes where an old
  producer fired, 0 regressions (0 shapes >2% slower). The old producers emit NO
  config on the other 12 shapes (cluster_m=1 decode etc.) -- the "A1 gap" this
  formula closes.
• geomean(seed / external ref) = 1.035 (curriculum), 1.083 (held-out fp16, 3 shapes
  with a ref) -- held-out >= curriculum, so no overfitting.
• accuracy 35/35 PASS (curriculum core; 48/48 incl. the FFI alt-seed arm; 4/4 fp16).
• FFI alt-seed KEPT: 3 curriculum shapes show a >2% win of the FFI explicit_epi_tile
  topology over the DEFAULT-layout core -- residual_add 4096^3 +54.0%,
  bias_residual_gelu 8192^3 +50.6%, silu 4096^3 +9.5% (the source-C/residual family
  where the core caps ab at 2 and the FFI ring recovers it). The autotuner keeps
  whichever of {core, FFI alt} wins.

[A] No-autotune default -- what a user gets with autotuning OFF (default_config()),
the promoted formula config vs the raw fragment default:

• 64x8192x8192 fp8 321.9 vs 29.3 TF/s (~11x); 512x4096x4096 bf16 697.2 vs 557.5;
  2048x4096x4096 bf16 1340.5 vs 1289.8. This is the primary user-visible win of the
  seed and does not depend on running the search.

[B] Effect on the autotune SEARCH (formula seed ON vs OFF, both on this full stack):

• Convergence: on 64x8192x8192 fp8 the seed lands the search AT the answer -- the
  first benchmarked config is ~1.05x the final best (seeded) vs ~4.77x (unseeded),
  reproducible across two reps. Both arms ultimately converge to the SAME winner
  ([64,64,128] cm1 ab12), re-benched cold-L2 to an identical 321.9 TF/s (any
  end-of-search delta in the tables is the autotuner's noisy in-process timer, not a
  real difference -- the cold-L2 re-bench is the trustworthy comparison).
• Ceiling: given PRs 1-3, the unseeded search on this stack already REACHES the good
  cluster_m=2 configs (e.g. 512x4096x4096 bf16 base and treat both reach
  [256,128,128] cm2 ab4 = 697.2). That reachability is PR3's contribution (widening
  the search space), which the seed-OFF arm confirms -- NOT a win attributable to
  this seed. The formula's marginal search-ceiling effect is small (a soft +3.9% on
  2048x4096x4096 bf16, where the seeded search reached the deep-AB [256,256,64] ab6
  tile in 1 of 2 reps and the unseeded did not). The seed's real value is the
  no-autotune default and convergence above, not moving the search ceiling.

Prerequisite edits bundled here (needed for the formula to emit valid configs):

• cute_mma.py: admit a rank-2 exact-shape residual aux on the explicit-epi-tile path
  (broadcast_axis in (1, None)) for the bias_residual_gelu / residual_add family --
  without it the FFI alt-seed's residual config is rejected (it drives the +50% wins).
• tcgen05_config.py: extend the 16-bit deep-AB (ab>3) SMEM-clamped admission from
  cluster_m=2 to cluster_m=1 DEFAULT layout, so the formula's bf16 decode tile
  ([64,32,256] ab4 = 196608 B) validates. max_ab_stages_that_fit still rejects the
  real cm1 256^2 overflow (clamped to ab<=3).
• _gelu_tanh_approx.py: lower cute.math.tanh with fastmath=True (hw tanh.approx.f32),
  recovering ~14% on the fused GEMM+GELU epilogue.

Also pins the promoted DEFAULT-layout seed's layout/ffi/epi-tile keys explicitly: on a
direct-entry-eligible shape the fragment defaults are FFI-biased, so a promoted seed
that omitted them merged into an invalid explicit_epi_tile+ffi hybrid and raised
InvalidConfig in the no-autotune / baseline compile path.

Known limitation, NOT fixed (out of scope): fp8 512x2048x2048 -- the formula emits
l2_groupings=[4] where the old producer emits [1]. Measured this run it is a near-tie
(seed/old 0.997), but the l2_groupings divergence is real; left for a follow-up.

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