Respect min and max of inputs to create more precise repro scripts

[crcrpar]

Related: #4529

When an nvfuser fusion definition takes signed int tensors as inputs, it's possible that those int tensors have a solid value range to sample from e.g. [0, vocab_size), for embedding, the FusionDefinition.last_repro_script seems to have lost such range information as it uses FusionDefinition.fake_inputs that is generated by fakifying the input tensors, leading to a loss of actual values.

Inspired by lightning-thunder's ExampleInputMetaData, I rather store the range info in a dataclass that can give us a string of make_tensor or equivalents to get tensors of the same value range.

For the same fusion definition as #4529, running the fusion definition as follows

inputs = [
    torch.randint(0, 10, (16,), dtype=torch.int32, device='cuda:0'),
    torch.testing.make_tensor((151936, 2048), dtype=torch.bfloat16, device='cuda:0'),
    torch.testing.make_tensor((2048,), dtype=torch.bfloat16, device='cuda:0'),
    torch.testing.make_tensor((2560, 2048), dtype=torch.bfloat16, device='cuda:0'),
    torch.testing.make_tensor((2560,), dtype=torch.bfloat16, device='cuda:0'),
]
fd.execute(inputs, save_repro_inputs=True, print_repro=True)

both fd.last_repro_script() and the fd.execute return

inputs = [
    torch.testing.make_tensor((16,), dtype=torch.int32, device="cuda:0", low=0, high=7,),
    torch.testing.make_tensor((151936, 2048), dtype=torch.bfloat16, device="cuda:0", low=-9.0, high=8.9375,),
    torch.testing.make_tensor((2048,), dtype=torch.bfloat16, device="cuda:0", low=-9.0, high=8.9375,),
    torch.testing.make_tensor((2560, 2048), dtype=torch.bfloat16, device="cuda:0", low=-9.0, high=8.9375,),
    torch.testing.make_tensor((2560,), dtype=torch.bfloat16, device="cuda:0", low=-9.0, high=8.9375,),
]

[Copilot]

Pull Request Overview

This PR improves the reproducibility of fusion definitions by preserving the original input range information instead of using generic fake inputs. The key changes include:

• Addition of the new dataclass InputReproducer to capture tensor metadata (min/max, shape, strides, etc.).
• Replacement of fake_inputs with last_input_reproducers in the execute and last_repro_script methods.
• Updated type annotations and diffs in the repro_script_for logic to utilize the new dataclass.

[Copilot]

Consider using isinstance(tensor, torch.Tensor) instead of comparing types directly to better support tensor subclasses.

Suggested change

	if type(tensor) is not torch.Tensor:
	if not isinstance(tensor, torch.Tensor):

[crcrpar]

Here I want to make sure tensor is compatible with torch.aminmax. What I really want to filter out is tensor subclasses that we cannot reconstruct with torch.testing.make_tensor, torch.rand, or torch.randint. Thus in short isinstance wouldn't be sufficing

[kiya00]

Do we also need additional information, such as requires_grad or storage_offset, for example? In Thunder, when this API is used for reporting, it also generates a corresponding Torch operator function for the NVFuser region. To ensure correctness, it constructs input tensors for this function that include properties like requires_grad and storage_offset (which were previously retrieved from FakeTensor).

[crcrpar]

I chose not to remove fake_inputs and to add last_input_reproducers with storage_offset and requires_grad