fix: support optional C in Gemm

[voltjia]

Summary

• Implement non-null optional C for the Gemm signature introduced by fix!: align Gemm public signature #878.
• Support independent and Y-broadcastable C across CPU, CUDA-family, Cambricon, Ascend, and PyTorch implementations.
• Keep the existing parameter matrix while changing it to independent C/Y, and add focused C=None and broadcast-C smoke coverage.

Motivation

#878 changes the public Gemm signature immediately and temporarily rejects non-null C. This follow-up completes the aligned contract without mixing the larger backend implementation into the API-only review.

Depends on #878.

Type of Change

• feat - new feature / new operator / new platform
• fix - bug fix
• perf - performance improvement (no behavioral change)
• refactor - code restructuring without behavior change
• test - adding or fixing tests only
• docs - documentation only
• build / ci - build system or CI configuration
• chore - tooling, formatting, or other non-code changes
• Breaking change (owned by fix!: align Gemm public signature #878)

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• Hygon (WITH_HYGON, PyTorch implementation)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Exact-head build, Ruff, and both clang-format checks passed for 256fcad. Platform unit workflows are not currently listed for this stacked-head update.

The implementation files remain aligned with the previously validated #870 implementation. The stack rewrite changes the parent interface commit and keeps all newly added test cases in this implementation PR.

Previous #870 implementation validation:
NVIDIA, Iluvatar, MetaX, Moore, Cambricon, Ascend - main CI passed
NVIDIA, Iluvatar, MetaX, Moore, Cambricon, Ascend - shadow CI passed
Current stacked head:
python -m py_compile tests/test_gemm.py scripts/run_host_overhead_control.py
passed
python scripts/generate_wrappers.py --devices cpu --ops gemm --strict-ops
passed
git diff --check b33fccd9..256fcadb
passed

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Exact-head platform job not listed; previous implementation passed	Native BLAS, cuBLASLt, and Torch paths
Iluvatar	Yes	Exact-head platform job not listed; previous implementation passed	Shared BLAS plus platform Add
MetaX	Yes	Exact-head platform job not listed; previous implementation passed	Shared BLAS plus platform Add
Cambricon	Yes	Exact-head platform job not listed; previous implementation passed	CNNL matmul plus op-tensor accumulation
Moore	Yes	Exact-head platform job not listed; previous implementation passed	Shared BLAS plus platform Add
Ascend	Yes	Exact-head platform job not listed; previous implementation passed	Native addmm/baddbmm input/output separation
Hygon	Yes	Not separately configured in this CI matrix	PyTorch implementation only
CPU	Yes	Exact-head build passed; previous implementation validated	Native and PyTorch implementations

Focused test notes

The inherited Gemm matrix now passes distinct C and Y tensors without adding
new parameter combinations. One smoke test covers both no-C public overloads,
and one float32 smoke test covers broadcast C.

Benchmark / Performance Impact

No benchmark was run. Independent or broadcast C requires a post-GEMM elementwise accumulation on classic BLAS and CNNL paths. Omitted C remains a single GEMM with beta zero.

Notes for Reviewers

• Review this PR relative to fix!: align Gemm public signature #878. The stacked diff is 13 implementation/test files and contains no documentation, example, or host-overhead script changes.
• The four platform Add includes are required because the shared BLAS template calls each platform's existing Add::Call specialization for non-null C. No Add implementation file is modified.
• Documentation remains unchanged: every current document and runnable example uses the still-valid three-tensor overload owned by fix!: align Gemm public signature #878.
• C may be independent and broadcastable to Y. C/Y aliasing is rejected because staged backends overwrite Y before the Add step.
• The existing Cartesian Gemm matrix now covers independent C/Y across its existing shapes, strides, dtypes, alpha/beta values, transposes, and implementations.
• The no-C overload and broadcast-C smoke tests live here rather than in the interface-only fix!: align Gemm public signature #878.
• The C=None smoke case verifies beta is ignored and an existing NaN-filled Y is not read.
• The broadcast smoke case computes its expected result before invoking Gemm, so accidental mutation of input C cannot affect the reference.
• No deprecated compatibility overload is retained.