[FEATURE REQUEST] `gemm` PyTorch kernel implementation

[LoserCheems]

Problem statement

The BLAS level-3 gemm kernel (general matrix-matrix multiply) is not implemented for the PyTorch backend in kernel_course.pytorch_ops. The README BLAS table lists no PyTorch support for gemm, even though GEMM is one of the most important operations in linear algebra and deep learning.

Without a PyTorch gemm kernel:

• there is no unified kernel API for $C = \alpha A B + \beta C$ on PyTorch tensors,
• cross-backend comparisons for GEMM are not possible within kernel-course,
• higher-level modules must rely directly on raw PyTorch matmul/linear primitives instead of a standard kernel abstraction.

Proposed solution

Add a PyTorch implementation of gemm under kernel_course.pytorch_ops that matches the Python reference semantics and follows existing backend patterns.

Concretely:

• Introduce kernel_course/pytorch_ops/gemm.py implementing $C = \alpha A B + \beta C$ for 2D torch.Tensor inputs.
• Align function signature and behaviour with other PyTorch kernels (argument order, dtype/device handling).
• Implement the operation using idiomatic PyTorch (e.g. torch.matmul or @ operator plus scaled add) while maintaining BLAS-like semantics.

Alternatives considered

Relying entirely on raw PyTorch matmul in higher-level code would:

• weaken the uniform kernel abstraction across backends,
• complicate benchmarking GEMM across Triton and CuTe,
• reduce the pedagogical clarity of mapping mathematical GEMM to backend implementations.

Implementation details

• Add kernel_course/pytorch_ops/gemm.py with a public gemm function.
• Accept scalars alpha, beta, matrices A, B, and C.
• Validate shapes and ensure compatibility with BLAS-style semantics.
• Update kernel_course/pytorch_ops/__init__.py to expose gemm.

Use case

The PyTorch gemm kernel will:

• provide a standard GEMM primitive through a unified kernel interface,
• act as a benchmark target for Triton and CuTe implementations,
• support construction of higher-level modules such as attention and MLP layers.

Related work

• Existing PyTorch kernels: pytorch_ops.copy, pytorch_ops.swap.
• PyTorch matmul and linear layer implementations.

Additional context

This issue contributes to completing the gemm row for the PyTorch column in the README BLAS table.