Refactor(linear): split LinearBackward kernel into 3 independent kernels

[chen2021673]

Summary

Architecture refactoring of Linear/Matmul/Outer kernels.

The core idea is separation of concerns — moving the decision of whether a gradient should be computed from the kernel layer up to the autograd layer, making kernels pure compute functions. At the same time, unified GEMM/SGEMV primitives are abstracted at the bottom layer to eliminate duplicated cuBLAS boilerplate.

Changes

• Autograd layer: LinearBackward and MatmulBackward are each decomposed into multiple independent Dispatcher calls. The needs_input_grad checks happen at the autograd layer, invoking only the kernels actually needed.
• Kernel layer: The monolithic LinearBackward is split into LinearBackwardInput / LinearBackwardWeight / LinearBackwardBias; MatmulBackward is split into MatmulBackwardInput / MatmulBackwardOther, with naming aligned to MatmulForward(input, other).
• File split: Matmul kernels are extracted from linear.cc / linear.cu into dedicated cpu/matmul.cc and cuda/matmul.cu, giving each file a single responsibility.
• GEMM primitive: New gemm.cuh / gemm.cu define the GemmParams struct and GemmCuda(), providing a unified wrapper over cublasGemmEx and cublasGemmStridedBatchedEx branching logic. GetCublasHandle() / GetCudaStream() are centrally defined and shared across linear.cu, matmul.cu, and outer.cu, eliminating duplicate definitions.
• SGEMV primitive: New SgemvParams struct and SgemvCuda() wrap the cublasSgemv call. LinearForward and LinearBackwardInput in linear.cu take the SGEMV path when bs==1 and fp32 (more efficient for matrix-vector shapes); bf16 falls back to GemmCuda since cublasSgemv does not support it. The fp32 backward path in outer.cu is migrated to SgemvCuda as well, eliminating inline cublasSgemv calls.