perf: Add native GPU implementation for GPUArray.transpose() and reshape()

[m96-chan]

Summary

GPUArray.transpose() and GPUArray.reshape() currently use CPU fallback (numpy) for all cases, which involves GPU→CPU→GPU data transfer overhead.

Current Implementation

def transpose(self, *axes: int) -> GPUArray:
    np_data = self.to_numpy()  # GPU→CPU
    result = np_data.transpose(*axes)
    return from_numpy(result.copy())  # CPU→GPU

Existing Native Implementations

• 2D transpose: ops.matmul.transpose() - native CUDA
• 3D (0,2,1): ops.tensor.transpose_3d_021() - native CUDA
• 4D transpose: Not implemented

Proposed Changes

1. GPUArray.transpose(): Use native implementations when available

   • 2D: delegate to matmul.transpose()
   • 3D with axes (0,2,1): delegate to transpose_3d_021()
   • Other cases: CPU fallback (or implement new kernels)

2. GPUArray.reshape():

   • For contiguous reshapes, use zero-copy view via narrow()
   • For non-contiguous, CPU fallback

3. New native kernels (optional):

   • transpose_4d_0213 for attention Q/K/V reshape
   • transpose_4d_0132 for K^T in attention
   • General transpose_nd kernel

Use Case

Whisper encoder/decoder attention uses 4D transposes heavily:

q = q.transpose(0, 2, 1, 3)  # [batch, seq, heads, dim] → [batch, heads, seq, dim]
k.transpose(0, 1, 3, 2)      # For K^T in attention scores

Priority

Medium - Current CPU fallback works but adds latency for real-time ASR.

[m96-chan]

Implemented in commit 51c1dfc:

Added native CUDA transpose kernels for common axis permutation patterns:

• 3D (0,2,1): transpose_3d_012 - swaps last two axes
• 4D (0,1,3,2): transpose_4d_0132 - swaps last two axes (K^T in attention)

GPUArray.transpose() now uses native GPU kernels for:

• 2D (1,0): matmul.transpose()
• 3D (1,0,2): tensor.transpose_3d_021()
• 3D (0,2,1): tensor.transpose_3d_012() [NEW]
• 4D (0,2,1,3): tensor.transpose_4d_0213()
• 4D (0,1,3,2): tensor.transpose_4d_0132() [NEW]

All tests pass (229 passed).