Forward strided-view matrix products to the storage-level mul!

[aalquwayfili]

Fixes #799.

Since #760, any product involving a strided GPU view goes to GPUArrays' generic matmul kernel. That skips the storage-level mul!(C, tA, tB, A, B, α, β) that back-ends overload to call their vendor libraries, so mul!(C, view(A, 1:4000, :), B) no longer uses cuBLAS. This PR forwards plain matrix-matrix products with views back to that method when every matrix is a valid BLAS operand:

• both wrapper chars are N, T or C;
• every matrix has a unit row stride and a leading dimension (stride(X, 2)) of at least its number of rows;
• the dimensions match.

Everything else stays on the generic kernels, as on main:

• Row-stepping views and reversed columns: the gemm wrappers in AMDGPU.jl and oneAPI.jl pass stride(X, 2) as lda without checking the layout. A row-stepping view gives wrong results there, and reversed columns make cuBLAS reject the call.
• Symmetric and Hermitian operands, and matrix-vector products: AMDGPU.jl's symm!, hemm! and gemv! only accept dense ROCArrays.
• Mismatched dimensions: Metal.jl's native mul! doesn't check them, so leaving those to the generic kernel keeps the DimensionMismatch error.

The #799 MWE on an RTX 5080 with CUDA.jl 6.4.2 drops from 2.5 ms (generic kernel) to 0.26 ms (cuBLAS). I measured that with the change applied to GPUArrays 11.5.16, since no CUDA.jl release supports GPUArrays 12 yet. In the same setup, linalg/mul!/strided-views passes 69/69 on CuArray. The new reversed-column tests error with CUBLAS_STATUS_INVALID_VALUE without the leading-dimension check. On this branch, JLArray/linalg passes on Julia 1.10, 1.12 and 1.13.

[aalquwayfili]

I narrowed this. The first version forwarded every product with a view, and I had only tested it with CUDA.jl. Reading the other back-ends, it would have broken them:

• AMDGPU.jl and oneAPI.jl pass any strided view to gemm! with lda = stride(A, 2) and no check on stride(A, 1), so view(A, 1:2:8, :) * B would have given wrong results.
• AMDGPU.jl's gemv!, symm! and hemm! only take ROCMatrix and ROCVector, so a view there would have been a MethodError.

Now only plain matrix-matrix products with contiguous columns are forwarded. The rest stays on the generic kernels, as on main. I updated the description.

The part about AMDGPU.jl and oneAPI.jl is from reading their code, I have no AMD or Intel GPU. Buildkite has not really run on this PR before: build 2011 finished in 3 seconds because I pushed while it was still a draft. So this push is the first run on the other back-ends.

If you want symm and gemv forwarded too, I can add an opt in for back-ends that handle views there.

[maleadt]

LGTM, thanks. Squashed, rebased, and slightly reworked.