Improve TorchAO quantization test coverage and XPU support

[jiqing-feng]

What does this PR do?

This PR improves the TorchAO quantization testing infrastructure with several fixes: enabling int4wo tests on Intel XPU, implementing _dequantize for TorchAO, fixing input dtype mismatches, and fixing training gradient underflow.

Changes

1. Enable int4wo tests on XPU: Removed the _int4wo_skip marker that restricted int4wo tests to CUDA only, allowing them to run on all accelerator backends.

2. XPU-specific int4 packing format: Added XPU-specific handling in _get_quant_config() — Intel XPU requires int4_packing_format="plain_int32" for Int4WeightOnlyConfig.

3. Fix input dtype casting: Introduced _get_dummy_inputs_for_model(model) helper in QuantizationTesterMixin to automatically cast floating-point input tensors to the model's parameter dtype, preventing dtype mismatches during quantized model inference.

4. Implement _dequantize for TorchAO: Added _dequantize() method in TorchAoHfQuantizer that iterates all nn.Linear modules, calls weight.dequantize() on TorchAOBaseTensor weights, and replaces them with standard nn.Parameter. Also fixed _verify_if_layer_quantized to check isinstance(module.weight, TorchAOBaseTensor) so dequantized layers are correctly detected as non-quantized.

5. Fix training gradient underflow: Changed autocast dtype from float16 to bfloat16 in _test_quantization_training. Float16's limited dynamic range (max ~65504, min subnormal ~5.96e-8) causes gradients to underflow to zero when passing through quantized tensor subclass operations; bfloat16 shares float32's exponent range and avoids this issue.

6. Reduce WanAnimate TorchAO test input sizes: Shrunk dummy inputs in TestWanAnimateTransformer3DTorchAo to avoid OOM on devices without FlashAttention (e.g. XPU, which falls back to math SDPA and materializes the full O(S²) attention matrix). Reduced hidden_states from (1,36,21,64,64) to (1,36,5,16,16) and face_pixel_values from (1,3,77,512,512) to (1,3,13,512,512), bringing self-attention sequence length from 21,504 to 320 and peak attention memory from ~74 GiB to ~16 MB. Face frame count (13) is chosen so the face encoder's two stride-2 convolutions produce temporal output 4, plus 1 padding = 5, matching hidden_states temporal dim.

[jiqing-feng]

Hi @sayakpaul . Would you please review this PR? Thanks!