Add fuse_matmul_add_bias_into_gemm_batched pass

[take-cheeze]

Implements the fuse_matmul_add_bias_into_gemm_batched pass from the docs/proposed_passes design draft. Split out of #4 so each pass is reviewed on its own.

Base branch: like #4, this is stacked on claude/initializers-non-constants-option-fun941 so its diff shows only the new pass. Retarget to main once that branch merges.

What it does

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. This means the pass does not change the default optimization pipeline's behaviour.

Tests

Adds unit tests — the fuse case, the dynamic-leading-dims case, and no-fuse guards (rank-2 MatMul, non-constant weight) — with onnxruntime numeric-equivalence checks where applicable.

---

Generated by Claude Code