Fix conv_transpose maxBufferLength failures on Metal via tiled unfold

[eyupcanakman]

Proposed changes

conv_transpose1d/2d/3d (and conv_general with input_dilation > 1) on Metal go through the explicit GEMM path, which allocates one full unfold buffer of implicit_M by implicit_K. For large outputs that buffer exceeds the device maxBufferLength and the allocation throws, even when the output itself fits in memory (#3082).

The explicit GEMM now unfolds and runs in row tiles that each stay under maxBufferLength, reusing a single unfold buffer so peak memory is one tile. Both the grouped and non-grouped paths are covered. The approach follows #3084.

MLX_CONV_UNFOLD_TILE_ROWS caps the tile size so the tiled path can be exercised on small inputs, which the added test uses.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)