[FEATURE REQUEST] `gemv` CuTe kernel implementation

[LoserCheems]

Problem statement

The BLAS level-2 gemv kernel has no CuTe backend implementation in this project. The README BLAS table shows an empty CuTe column for gemv, which prevents full cross-backend coverage for matrix-vector multiply.

Without a CuTe gemv kernel:

• users cannot see how GEMV is expressed with CuTe’s layout and tiling abstractions,
• there is no CuTe performance baseline to compare against PyTorch and Triton gemv,
• CuTe-based higher-level modules lack a fundamental building block.

Proposed solution

Implement a CuTe-based gemv kernel matching the Python reference semantics and aligning with the project’s backend structure.

Concretely:

• Add a CuTe gemv kernel in the appropriate CuTe backend directory, implementing $y = \alpha A x + \beta y$.
• Use CuTe primitives to describe the matrix layout, vector access, and thread scheduling.
• Ensure API parity with other backends so callers can dispatch to CuTe gemv uniformly.

Alternatives considered

Alternatives such as omitting CuTe gemv or relying solely on other backends would:

• reduce the educational impact of comparing CuTe to PyTorch/Triton on GEMV,
• leave the CuTe column partially empty in the README BLAS table,
• limit the ability to build CuTe-based end-to-end examples.

Implementation details

• Establish the file layout and build integration for CuTe kernels.
• Implement gemv using CuTe constructs for row-major/column-major layouts and tiling.
• Match numerical behaviour and broadcasting of scalars alpha and beta with the Python reference.
• Integrate with planned testing and benchmarking for gemv.

Use case

The CuTe gemv kernel will:

• demonstrate GEMV implementation in CuTe,
• enable detailed performance comparisons across backends,
• support more complex CuTe-based BLAS and Transformer kernels.

Related work

• CuTe/CUTLASS examples of GEMV/GEMM.
• Standard BLAS gemv implementations.

Additional context

This issue complements the gemv Python/PyTorch/Triton feature requests and helps complete the CuTe column in the README BLAS table.