Add fuse_consecutive_mul and batched MatMul+Add->Gemm passes

[take-cheeze]

Implements two passes from the docs/proposed_passes design drafts.

Base branch: this PR is stacked on claude/initializers-non-constants-option-fun941 so its diff shows only the two new passes. onnxsim's submodule (and its SetInitializersAsConstants usage) already depends on that branch's commit, so basing here keeps the submodule bump non-regressing. Once that branch merges, retarget this PR to main.

fuse_consecutive_mul (default Fuse pass)

Folds Mul(Mul(X, C1), C2) -> Mul(X, C1 * C2) when C1/C2 are constants (initializers or Constant nodes) and the inner Mul feeds only the outer one. The combined scale is materialised with numpy-style broadcasting, so the rewrite is numerically identical (X * C1 * C2 at every position). Covers e.g. PoolFormer's LayerScale, exported as a per-channel (C,1,1) scale times a scalar factor. Conservatively limited to FLOAT/DOUBLE constants; other dtypes are left to constant folding.

fuse_matmul_add_bias_into_gemm_batched (PassType::Other, opt-in)

Rewrites batched MatMul(X[>=3D], W[K,N]) + b into:

X2 = Reshape(X, [-1, K])
G  = Gemm(X2, W, b)      # alpha=beta=1, transA=transB=0
A  = Reshape(G, [d0, ..., d_{r-2}, N])

extending the 2-D-only fuse_matmul_add_bias_into_gemm to the rank-≥3 activations transformer linear layers use. Static leading dims emit a constant output shape; dynamic leading dims rebuild it via Shape/Slice/Concat (requires opset ≥ 10). W must be a 2-D constant and b a 1-D bias broadcastable over the N axis.

Registered as PassType::Other so it stays out of the default fuse set (GetFuseAndEliminationPass) — it is a graph-shape rewrite that trades a batched MatMul for Reshape + Gemm + Reshape and is not guaranteed to be faster, so it is invoked explicitly by name.

Tests

Adds unit tests for both passes — fuse cases, no-fuse guards (multiple-use, non-constant operand, rank-2, non-constant weight), per-channel and dynamic-shape cases — with onnxruntime numeric-equivalence checks. Full suite: 175 passed, 5 skipped (torchvision, torch not installed), 6 pre-existing xfails.

---

Generated by Claude Code

[take-cheeze]

Superseded by #5 (fuse_consecutive_mul) and #6 (fuse_matmul_add_bias_into_gemm_batched) — the two passes are now split into one PR each so they can be reviewed and merged independently. Closing this combined PR.

Note: I'm intentionally not deleting the claude/onnx-optimizer-passes-tx9y3n branch. onnxsim/onnxsim's submodule was already bumped to this branch's tip (c9a482c, which contains both passes) in onnxsim/onnxsim#523, so the commit needs to stay reachable until that submodule is repointed at the merged commits.

---

Generated by Claude Code