feat: route infiniop gemm through InfiniOps

[voltjia]

Summary

This PR routes the existing infiniopGemm C API through the InfiniOps C++ operator API.

It is stacked on #1198 so the diff only contains the GEMM integration changes on top of submodules/InfiniOps.

Changes

• Replace the infiniop GEMM descriptor/dispatch implementation with an adapter that calls infini::ops::Operator<infini::ops::Gemm>::Call(...)
• Link infiniop against submodules/InfiniOps/build/src/libinfiniops.so
• Add InfiniOps include/rpath wiring in xmake
• Add CUDA 13 target include fallback for NVIDIA builds
• Advance submodules/InfiniOps from 316ad6b0 to 9444f9c3 because 316ad6b0 fails to build libinfiniops.so with generated optional-parameter call instantiations

Verification

Verified on nvidia:

cmake -S submodules/InfiniOps -B submodules/InfiniOps/build \
  -DPython_EXECUTABLE=/home/huangjiacheng/.venv/bin/python \
  -DCMAKE_CUDA_COMPILER=/usr/local/cuda-13.0/bin/nvcc \
  -DWITH_CPU=ON -DWITH_NVIDIA=ON -DWITH_TORCH=ON \
  -DGENERATE_CPP_OPERATOR_API=OFF -DGENERATE_PYTHON_BINDINGS=OFF
cmake --build submodules/InfiniOps/build --target infiniops --parallel 8

PATH=/usr/local/cuda-13.0/bin:/home/huangjiacheng/.local/bin:$PATH \
  CUDA_HOME=/usr/local/cuda-13.0 \
  ~/.local/bin/xmake f -y --require=no --nv-gpu=y --cpu=y --cuda_arch=sm_80 --cu=/usr/local/cuda-13.0/bin/nvcc

PATH=/usr/local/cuda-13.0/bin:/home/huangjiacheng/.local/bin:$PATH \
  CUDA_HOME=/usr/local/cuda-13.0 \
  ~/.local/bin/xmake build --jobs=1 infiniop

Result:

build ok, spent 2.738s

Runtime smoke:

INFINI_ROOT=/tmp/infinicore-gemm-install \
LD_LIBRARY_PATH=/tmp/infinicore-gemm-install/lib:/tmp/infinicore-infiniops-gemm-clean/submodules/InfiniOps/build/src:/home/huangjiacheng/.venv/lib/python3.10/site-packages/torch/lib:$LD_LIBRARY_PATH \
/home/huangjiacheng/.venv/bin/python test/infiniop/gemm.py --nvidia --debug

Result:

Test passed!

Also verified ldd build/linux/x86_64/release/libinfiniop.so resolves libinfiniops.so from submodules/InfiniOps/build/src/libinfiniops.so.