Add MobileNet/ResNet8 layer-accurate ConvGrad kernel regression tests

15 single-op ConvGradX/W test directories that mirror actual backward
shapes in MobileNetV1 0.25x (stem conv, DW/PW blocks 0/1/3/7/11 incl.
stride-2 variants) and ResNet8 layer3_conv2. Each has inputs.npz
(random seeded), outputs.npz (PyTorch `torch.nn.grad.conv2d_{input,
weight}` reference), and single-op ONNX.

Post the parent commit (ConvGradX global-pad + ConvGradW accum fix),
13/15 pass bit-exact:
- ConvGradX_DW_block_0 / ConvGradW_DW_block_0: tiler "single minimal
  outer shape" legalization failure in standalone mode only; the same
  shape tiles fine inside the full MobileNet training pipeline.
- ConvGradW_R8_L3_conv2: fails 100% due to the mixed C_out+HW tile
  memset bug (same L1 dW slice re-memset'd across HW subtiles). ResNet8
  training doesn't actually hit this case, so it's parked as a
  known-issue regression test.

The debug/gen_kernel_tests.py generator builds all shapes via
torch.nn.grad APIs for reproducibility.

Author: runwangdl