[FEATURE REQUEST] `scal` CuTe kernel implementation

[LoserCheems]

Problem statement

The BLAS level-1 scal kernel (scale a vector in-place) does not yet have a CuTe (CUTLASS/CuTe-style) backend implementation in this project. While the README BLAS table lists a CuTe column, the scal row is currently unimplemented, preventing a complete set of cross-backend examples for this fundamental operation.

Without a CuTe scal kernel:

• users cannot study how a simple BLAS-1 operation maps onto CuTe primitives and memory layouts,
• there is no CuTe baseline for performance comparison against PyTorch and Triton scal implementations,
• higher-level modules that aim to mix and match backends cannot rely on a fully populated BLAS-1 set.

Proposed solution

Add a CuTe backend implementation of the scal kernel that follows the mathematical semantics of the Python reference and matches the public API conventions used for other backends.

Concretely:

• Introduce a CuTe-based scal kernel (file and namespace layout to match existing CuTe kernels once they are introduced in the repository).
• Implement $y = \alpha y$ for 1D vectors using CuTe primitives and recommended memory layout abstractions.
• Align the entry-point interface with the other backends so that higher-level code can dispatch to CuTe scal in a uniform way.

Alternatives considered

Alternatives such as skipping scal for CuTe or relying on other backends for performance analysis would:

• reduce the educational value of comparing CuTe against PyTorch/Triton on a simple kernel,
• leave the CuTe column in the README BLAS table partially incomplete,
• make it harder to build CuTe-based examples that mirror the Python/PyTorch/Triton stacks.

Implementation details

• Define the precise file path and build integration for CuTe kernels (e.g. a cute_ops or analogous directory, aligned with project conventions).
• Implement the scal kernel using CuTe constructs appropriate for vector operations, focusing on clear mapping between math and low-level execution.
• Ensure numerical equivalence with the Python scal reference across supported dtypes.
• Integrate the CuTe scal implementation into any existing or planned dispatch, testing, and benchmarking infrastructure.

Use case

The CuTe scal kernel will:

• demonstrate how a simple BLAS-1 operation is expressed in CuTe,
• serve as a stepping stone toward more complex CuTe-based BLAS and Transformer modules,
• provide a basis for cross-backend performance benchmarking and tuning.

Related work

• BLAS level-1 scal implementations in other GPU libraries.
• CuTe/CUTLASS examples for vector and memory-bound kernels.

Additional context

This issue fits into a broader effort to provide CuTe counterparts for each BLAS kernel in the README, starting from the simplest operations (copy, swap, scal).