issue/1465 fix: adapt Gemm call to optional c API

[voltjia]

What

• Pass an empty std::optional<infini::rt::TensorView> for the new Gemm input-C parameter.
• Advance the InfiniOps submodule from 8348acc69fd0c792f7243bbc7404dd8db4205798 to the focused prerequisite 1b9c37659211eeaeedb48b9691f79e4c9e4412f9 (InfiniOps #878).

Why

InfiniOps #878 split Gemm's optional input C from output Y. InfiniCore still instantiated the old seven-argument call, so the shared library retained an unresolved old-signature symbol and Python import failed at dlopen when used with current InfiniOps.

Scope

This restores the C=None / effective beta=0 path used by current inference and Matmul. It does not add non-null C or general beta != 0 support; that implementation is still tracked by the draft InfiniOps #870.

Screenshots: N/A (backend API compatibility only).

Validation

All build and runtime validation ran on ssh nvidia in accelerator-dev/nvidia:latest (sha256:dd94fce2f83a180e2271f02f69713de5360aaf916b01e1f7f54173519fd54efd) on NVIDIA A100 GPUs.

• PR dependency pin (InfiniOps@1b9c3765): InfiniOps and InfiniCore build/install, editable Python install, import infinicore, and Qwen3-0.6B CUDA inference passed (42,515.44 ms).
• Focused Matmul regression: 42/42 cases passed across seven shape/stride groups, FP16/BF16/FP32, and out-of-place/in-place output paths.
• Current InfiniOps master (e733e325): rebuild/install/import and Qwen3-0.6B CUDA inference passed (42,494.8 ms). The first inference attempt selected GPU 0 while an unrelated VLLM process held 75,558 MiB and failed allocation; the identical inference rerun on free GPU 1 passed.
• python scripts/format.py --ref origin/main --path src/infinicore/ops/gemm/gemm_infiniops.cc --check
• git diff --check

Fixes #1465