nvfp4tensor: improve printing

Summary:

Makes printing of linears with NVFP4 weights more descriptive, such as

```python
(gate_proj): Linear(in_features=2048, out_features=1408, weight=NVFP4Tensor(self._is_swizzled_scales=True, self.use_triton_kernel=False, self.act_quant_kwargs=None))
```

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

ghstack-source-id: b012e61
ghstack-comment-id: 3340883530
Pull Request resolved: #3086

Author: vkuzo

torchao/prototype/mx_formats/nvfp4_tensor.py

@@ -133,6 +133,9 @@ def __new__(
     def __repr__(self):
         return f"NVFP4Tensor: blockwise_scales: {self._scale_e4m3}, per_tensor_scale: {self._per_tensor_scale}, d: {self.qdata}, d_hp: {self.to_dtype(self._orig_dtype)}"
 
+    def _quantization_type(self):
+        return f"{self._is_swizzled_scales=}, {self.use_triton_kernel=}, {self.act_quant_kwargs=}"
+
     @classmethod
     def __torch_dispatch__(cls, func, types, args, kwargs=None):
         # Use NVFP4-specific ops table