Test failure if we use torch 2.5

[liu-jc]

Is your feature request related to a problem? Please describe.
If we don't require torch version < 2.5, the build test will fail, particularly test/module/test_attention.py::test_all_masked:

uni2ts/test/module/test_attention.py

Line 85 in 2ba614d

def test_all_masked(

Describe the solution you'd like
The reason might be that torch 2.5 updates the function nn.functional.scaled_dot_product_attention (see here https://github.com/pytorch/pytorch/releases/tag/v2.5.0)
It mentions:

Before this PR, fully masked rows in the attn_mask passed to nn.functional.scaled_dot_product_attention would yield NANs in the output, after this PR, fully masked rows yield 0s.

I think later maybe we would like to change the test function to assert all 0s.
But I am not sure if we have special usage for torch.nan in output (like if we change to 0s for fully masked rows, it will break the pipeline)
@chenghaoliu89 @cuthalionn Any comments/suggestions?