Update on "Improve QAT nvfp4 numerics"

**Summary:** Similar to #2986,
this commit improves the prepare vs convert SQNR of NVFP4 QAT
from 12 to 36 with `use_per_tensor_scale`, and 12 to inf without.
This is achieved by mimicking the PTQ flow more closely,
in particular, in descending order of significance:

1. Simulate `f4_unpacked_to_f32` and `f32_to_f4_unpacked`,
   but in `torch.int32` instead of `torch.uint8`
2. Do not cast intermediate fake quantized values to original
   dtype, e.g. bf16 which loses some fidelity from fp32
3. Fake round blockwise scales to float8

**Test Plan:**
```
python test/quantization/test_qat.py -k test_qat_nvfp4
python test/quantization/test_qat.py -k test_quantize_api_nvfp4
```

End-to-end tests TBD.

[ghstack-poisoned]

Author: andrewor14

test/quantization/test_qat.py

@@ -2096,6 +2096,7 @@ def test_quantize_api_nvfp4(self, use_per_tensor_scale: bool):
             target_convert_sqnr=float("inf"),
         )
 
+    @unittest.skipIf(not is_sm_at_least_89(), "Need sm89+")
     @unittest.skipIf(not _CUDA_IS_AVAILABLE, "skipping when cuda is not available")
     @parametrize("use_per_tensor_scale", [True, False])
     def test_qat_nvfp4(self, use_per_tensor_scale: bool):