Generalized cudax::copy_bytes for mdspan

[fbusato]

closes #7554

Description

cuda::copy_bytes(mspan) is currently limited to same layout, contigous mdspan.

This PR extends copy_bytes for mdspan-based host/device transfers so it works correctly and efficiently across contiguous and strided layouts, including different rank and mixed layout/order cases.
Common use cases are: padded tensors, column-major views, subviews, and other layout_stride mappings.

Main features:

• Optimal (lowest) number of transfers.
• Compatibily with any strided layout, including edge cases, such as transposition.
• Use cuMemcpyBatchAsync on CTK 13+.

The main challenge is use two different stride orderings for different purposes:

• Stride-sorted order: finding common contiguous regions / tile size.
• Original order: required for correct logical index mapping when source and destination stride permutations differ.

Details:

• If source/destination have the same stride order, iterating in stride-sorted order is safe and efficient.
• If stride orders differ, using sorted order for address generation can mismatch source/destination linearization. In that case, iterator addressing switches to original tensor order, while still using the computed tile size from layout analysis. This preserves correctness and still exploits contiguity where available.

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 2h 17m: Pass: 100%/137 | Total: 3d 01h | Max: 1h 53m | Hits: 90%/279164

See results here.