Metal apple silicon fixes

[tyt2y3]

Two independent Apple-Silicon Metal fixes, one commit each:

1. MetalDevice::new panics on Apple Silicon

MetalDevice::new does Device::all().swap_remove(ordinal). On Apple Silicon, MTLCopyAllDevices returns no devices, so swap_remove(0) panics on an empty vec — Metal is unusable on M-series:

thread 'main' panicked: swap_remove index (is 0) should be < len (is 0)

Fix: use Device::system_default() (MTLCreateSystemDefaultDevice) for the default device, falling back to all() with a bounds check for explicit additional-GPU ordinals.

2. conv weight-gradient is wrong on Metal — training silently diverges

The conv1d / conv2d / conv_transpose2d backward computes the kernel gradient by convolving a transposed (non-contiguous) input:

let grad_kernel = arg.transpose(0, 1)?.conv2d(&grad.transpose(0, 1)?, ...)?...

The Metal im2col reads its input as if contiguous, so the weight gradient is incorrect on Metal. The forward pass and the input gradient are correct — so inference works fine, but training diverges: every conv kernel is updated in the wrong direction and the loss climbs.

Fix: force the transposed operands contiguous before the conv. No-op on CPU.

Verification

Minimal CPU-vs-Metal check of conv2d through loss.backward() (max abs diff):

	forward	input-grad	weight-grad
before	2e-7	2e-6	2.95e1
after	2e-7	2e-6	7.6e-6

A small residual U-Net that diverged when trained on Metal (loss climbing, PSNR going negative) trains normally after this change, matching the CPU result step-for-step. group_norm, reshape/permute, and the conv input gradient were all already correct on Metal — only the conv weight gradient was affected.