Handle non-multiple-of-8 spatial dims in depthwise conv2d Metal path

[qflen]

Summary

Fixes #3324. Relaxes the oS % 8 == 0 dispatch gate on the depthwise Metal kernel and makes it tail-aware, so non-mod-8 output spatial dims stay on the fast path instead of falling back to the gemm path.

Root cause

depthwise_conv_2D_gpu dispatches (tc=8, tw=8, th=4) threadgroups across (C/tc, oW/tw, (oH/th)*N). In dispatch_conv_2D_gpu the selection guard is

conv_params.oS[0] % 8 == 0 && conv_params.oS[1] % 8 == 0

so any non-mod-8 output spatial shape routes to implicit_gemm_conv_2D_gpu / explicit_gemm_conv_group_ND_gpu, which is 2-5x slower for depthwise on M-series. The issue title says "mod 16" but the real boundary is mod 8: 240 (not mod 16, but mod 8) stays on the fast path, while 60, 30, 15 do not.

Change

• mlx/backend/metal/conv.cpp
  • depthwise_conv_2D_gpu: ceil-div grid_dims for the two spatial dims so tail tiles are dispatched.
  • dispatch_conv_2D_gpu: drop the oS % 8 == 0 guard.
• mlx/backend/metal/kernels/conv.metal
  • depthwise_conv_2d: ceil-div n_tgblocks_h to match the dispatch, and early-return from threads whose oh / ow are past oS before the output store. The shared-memory input load is already bounds-checked (ih/iw guards), and out-of-range threads still cooperate in that load, so only the final store needs guarding.

The C, kernel-size, stride, and wt_strides constraints on the depthwise path are unchanged.

Perf (M5 32GB, issue reproducer, 3x3 / pad=1 / stride=1 / depthwise, median of 3 runs)

channels	res	baseline ms	patched ms	speedup
384	60x60	1.64	0.34	4.8x
768	30x30	0.94	0.28	3.4x
1536	15x15	0.61	0.23	2.7x

Mod-8 shapes are unchanged (within +/-2% across runs). The previously-slow shapes now match their mod-8 siblings: 60x60 @ 384ch is 0.34 ms vs 64x64 @ 384ch at 0.35 ms.

Tests

• Added test gpu depthwise conv2d non-mod-8 spatial in tests/gpu_tests.cpp. It computes depthwise conv2d on GPU and CPU for six shapes (issue reproducer sizes, non-square, asymmetric spatial alignment, stride 2) and asserts allclose at 1e-4.
• Full ./build/tests/tests suite: 245 passed, 0 failed.

Checklist

• Ran clang-format on all changed files (no changes).
• Added a test (tests/gpu_tests.cpp).
• Benchmarked before / after, numbers above.
• No API changes; no doc changes needed.