rocBLAS updates for 1.13

[luraess]

No description provided.

[luraess]

The codegen failures for 1.13 relate to fp atomics issue not related to this PR.

[luraess]

Fix scalar indexing in strided Float16 matmul on Julia 1.13. Julia 1.13 changed the matmul dispatch chain so that types not covered by BLAS fall through to _generic_matmatmul_nonadjtrans!, which does scalar indexing. AMDGPU's generic_matmatmul_wrapper! override only matched T<:ROCBLASFloat, missing Float16. Extending it to T<:ROCBLASFloatWithHalf routes Float16 through the generic_matmatmul! override, which correctly falls back to GPUArrays.generic_matmatmul!.