[FEATURE REQUEST] `gemm` Triton kernel implementation

[LoserCheems]

Problem statement

The BLAS level-3 gemm kernel has no Triton implementation in kernel_course.triton_ops. The README BLAS table shows an empty Triton column for gemm, even though GEMM is the canonical example for high-performance GPU kernels.

Without a Triton gemm kernel:

• users cannot study or experiment with GEMM implementations in Triton within this project,
• there is no Triton baseline for performance comparison against PyTorch and CuTe GEMM,
• the Triton backend lacks a flagship example that ties together tiling, memory layouts, and parallelism.

Proposed solution

Add a Triton implementation of gemm under kernel_course.triton_ops that matches the Python/PyTorch semantics and showcases best practices for GEMM-style kernels.

Concretely:

• Introduce kernel_course/triton_ops/gemm.py defining a Triton JIT GEMM kernel and Python wrapper.
• Implement tiled matrix multiplication with configurable block sizes and good memory access patterns.
• Ensure numerical equivalence with the Python reference across supported dtypes.

Alternatives considered

Relying solely on PyTorch or CuTe for GEMM would:

• miss a key educational opportunity to present GEMM in Triton,
• limit performance and implementation comparisons across backends,
• leave the Triton column incomplete in the README BLAS table.

Implementation details

• Add kernel_course/triton_ops/gemm.py containing:
  • a Triton kernel with parameters for tiles of A, B, and C,
  • a wrapper to configure grid/block sizes and launch the kernel.
• Handle arbitrary matrix shapes with proper bounds checks.
• Consider multiple dtypes (e.g. float16, float32) and accumulation strategies.
• Integrate with future tests and benchmarks.

Use case

The Triton gemm kernel will:

• act as a central example of high-performance GPU kernel development in this project,
• enable detailed performance comparisons with PyTorch and CuTe GEMM implementations,
• serve as a building block for Transformer modules and other complex workloads.

Related work

• Existing Triton kernels: triton_ops.copy, triton_ops.swap.
• Triton GEMM tutorials and reference implementations.

Additional context

This issue is a key part of extending the Triton backend to cover BLAS level-3 operations listed in the README, starting with gemm.