feat: support tile-misaligned GPTQ Marlin shapes

[ZX-ModelCloud]

Summary

• Allow explicitly selected GPTQ Marlin layers to handle packed K/N shapes that do not already fill a 64 x 128 thread tile.
• Zero-pad packed weights, scales, and bias during post_init, then pad activations and slice outputs around the existing Marlin GEMM.
• Keep aligned layers on the original call path and keep runtime padding opt-in for automatic backend selection.
• Reject partially packed dimensions and unsupported activation-order tails before kernel launch.

Compatibility

• Supports symmetric GPTQ 4-bit and 8-bit weights with FP16 or BF16 compute.
• Supports channelwise quantization expressed as either group_size=-1 or group_size=K.
• Does not change CUDA sources, operator registration, or the aligned-shape execution path.
• Tile-misaligned explicit Marlin requests may trade performance for shape compatibility; automatic selection remains conservative.

Tests

• ruff check gptqmodel/nn_modules/qlinear/marlin.py gptqmodel/utils/marlin.py tests/test_marlin_jit.py
• pytest -q tests/test_marlin_jit.py -m 'not cuda' (33 passed)
• pytest -q tests/test_marlin_jit.py::test_marlin_cuda_padded_shape_matches_dequantized_reference (16 passed on sm80)