Extend NVFP4 QAT to support emulation

[andrewor14]

As of #3050, NVFP4 QAT is only usable on Blackwell GPUs. This is because the forward pass reuses all the quantization primitives used in NVFP4Tensor in the PTQ path. However, a potentially important use case is to train the model on other GPU architectures (e.g. A100 or H100) using QAT, save the model, and load it on a different machine with B200 support for inference.

More details: The main known numerical discrepancy comes from these two paths:

# (1) Current implementation: only supported on B200 currently
# https://github.com/pytorch/ao/blob/main/torchao/prototype/qat/nvfp4.py#L82
_addmm_nvfp4_dispatch(nvfp4_input, nvfp4_weight.t(), None, bias)

# (2) More general implementation: probably works on A100 or H100, need to verify
# https://github.com/pytorch/ao/blob/main/torchao/quantization/qat/linear.py#L115
orig_dtype = torch.bfloat16
fq_input = nvfp4_input.to_dtype(orig_dtype)
fq_weight = nvfp4_weight.to_dtype(orig_dtype)
F.linear(fq_input, fq_weight, bias)

(1) and (2) don't match currently. (1) gives inf SQNR compared to the PTQ flow, while (2) gives only ~21 for this QAT test. We want to restructure our code to (2) so we can reuse existing QAT components like FakeQuantizer, but keep the inf SQNR from (1).