Storage-level mul! gives wrong results or MethodErrors for SubArray operands #1161

The storage-level mul! methods in src/blas/highlevel.jl pass SubArrays of ROCArrays to rocBLAS without checking their layout, and some of the wrappers they call don't accept views at all:

using AMDGPU, LinearAlgebra
A = ROCArray(rand(Float32, 8, 4)); B = ROCArray(rand(Float32, 4, 5)); x = ROCArray(rand(Float32, 4))
S = ROCArray(rand(Float32, 8, 8))

# row-stepping view: silently wrong result (lda = stride(A, 2) and stride(A, 1) is ignored)
C = mul!(ROCArray(zeros(Float32, 4, 5)), 'N', 'N', view(A, 1:2:8, :), B, true, false)
Array(C) ≈ Array(A)[1:2:8, :] * Array(B)   # false

# reversed columns: ROCBLASError(rocblas_status_invalid_size)
mul!(ROCArray(zeros(Float32, 8, 5)), 'N', 'N', view(A, :, 4:-1:1), B, true, false)

# any view, even a contiguous one, on the gemv/symv/hemv/symm/hemm paths: MethodError
mul!(ROCArray(zeros(Float32, 4)), 'N', view(A, 1:4, :), x, true, false)       # no method matching gemv!
mul!(ROCArray(zeros(Float32, 4, 5)), 'S', 'N', view(S, 1:4, 1:4), B, true, false)   # no method matching symm!

The mul! methods take StridedROCMatrix/StridedROCVector. gemm! and gemm_ex! take strided arrays too, but compute lda = max(1, stride(A, 2)) without checking stride(A, 1). gemv!, symv!, hemv!, symm! and hemm! (src/blas/wrappers.jl) only accept dense ROCMatrix/ROCVector. The 'S'/'H' branch also has no check that eltype(A) == eltype(B) == eltype(C), so mixed element types end up in a MethodError too.

The public mul!(C, A, B) isn't affected, because GPUArrays intercepts products with views before they reach these methods. JuliaGPU/GPUArrays.jl#800 starts forwarding matrix-matrix products with views here again, but only N/T/C products where every matrix has a unit row stride and stride(X, 2) >= size(X, 1). It keeps Symmetric/Hermitian and matrix-vector products away because of the MethodErrors above. Calling the storage-level mul! directly is public API since Julia 1.13 (JuliaLang/LinearAlgebra.jl#1671), though. CUDA.jl added the equivalent guards in JuliaGPU/CUDA.jl#3316:

• only call rocBLAS when every matrix has stride(X, 1) == 1 and stride(X, 2) >= max(1, size(X, 1)), and every vector has a positive stride;
• check that element types match on the symm/hemm branch;
• widen the BLAS 2 and symm/hemm wrappers to strided arrays, or fall back for views;
• send everything else to GPUArrays.generic_matmatmul!.

With those in place, GPUArrays could forward Symmetric, Hermitian and matrix-vector products with views as well.

Reproduced with AMDGPU.jl 2.9.1 and GPUArrays 12.1 on a gfx1036 iGPU, Julia 1.12.7.