Add explicit tests + benchmark shapes for copy_transpose

[debashishc]

Summary

Add explicit correctness and benchmark shape grids for copy_transpose, including tail/tile edge cases.

Motivation / Use Case

We need repeatable shape coverage and tail handling checks for tile sizes 16 and 32.

Proposed Solution

• Correctness tests (tile_size 16 and 32):
  • Square: 512x512, 1024x1024
  • Rectangular: 256x1024, 1024x256
  • Edge: 255x257, 257x255
• Benchmarks:
  • Square: 1024x1024, 2048x2048, 4096x4096
  • Rectangular: 512x4096, 4096x512
• If alignment constraints exist, document them and add tests that verify behavior.

Scope Alignment

v0.1 scope (Weeks 0-2)

Alternatives Considered

Only test square matrices or single tile size.

Additional Context

This is the baseline kernel; tests must be fast and stable.