[Feat][Ops]: BmmFwdOp: manifest + kernel/op + tests + bench (fp16/bf16/fp8)

[RuneFang]

Parent Issue

Sub-issue of #400.

I'd be happy to implement the next two operators.

Task Type

• Manifest / Spec (required for any public Op)
• Kernel / Op Implementation
• Correctness Tests / Workloads
• Benchmark / Performance

Description

Track the BMM slice of #400. The 2 nominal ops listed there (bmm_fp16, bmm_fp8) map to 2 TileOPs ops, following the same input-inferred / call-time-JIT design as GemmOp:

nominal op	Realized as	Scope
bmm_fp16	BmmFwdOp (fp16 + bf16)	this issue
bmm_fp8	BmmFp8Op (fp8, epilogue-scale + block-scaled paths)	follow-up sub-issue

Semantics (settled):

• Strict torch.bmm contract — 3D×3D only, a=[B, M, K], b=[B, K, N] → d=[B, M, N]. No batch-dim broadcasting, no matmul-style shape promotion.
• Transpose-free primary layout matching torch.bmm inputs; NT / other layouts are separate manifest entries if demand appears.
• Fully input-inferred: batch, m, n, k, and dtype are read from a / b at forward() time; nothing pinned at construction. Roofline binds these dims via func mode.
• Reference API: torch.bmm. Dtype coverage: float16, bfloat16. Accumulation in float32.

Checklist

• Public Op has a tileops/manifest/ entry (or updates an existing entry).
• Manifest signature declares inputs, outputs, params, shape rules, and dtype coverage.
• Manifest workloads declare benchmark shapes/dtypes; unit-test edge cases are not generated from manifest workloads.
• Manifest roofline is present and consumable by op.eval_roofline().
• Manifest source declares kernel, op, test, bench, and source.kernel_map when dispatching kernels.
• Op constructor, forward(), and default_kernel_map match the manifest entry.
• Tests use an independent reference implementation and cover relevant FP16/BF16 and edge cases.
• Benchmarks consume manifest workloads and record at least one non-tileops baseline unless explicitly justified.
• PR title and commits follow the current [Feat][Scope] / [Perf][Scope] / [Fix][Scope] convention.
• Implementation follows Google Python Style for code and docstrings.

[lcy-seso]

Thanks for your interest in TileOps! Looking forward to your PR. 😊