[bugfix] fix flex-attention not supported on Ascend NPU, update BlockMask type annotations to str

[FightingZhen]

What does this PR do?

1. Problem 1
Flex-attention has not been fully verified on Ascend NPU yet. In PR #37866 , the func is_torch_flex_attn_available defined in below code does not contain logical judgment about Ascend NPU. In this situation, when we use torch>=2.5.0, this func will return True on Ascend NPU, which is not correct.

transformers/src/transformers/utils/import_utils.py

Line 412 in d03a3ca

def is_torch_flex_attn_available():

2. Problem 2
If func is_torch_flex_attn_available return False on Ascend NPU as expected, object BlockMask will not be imported in below code

transformers/src/transformers/masking_utils.py

Line 29 in a5a0c7b

from torch.nn.attention.flex_attention import BlockMask, create_block_mask

But this object is directly called in type annotations in below code now, directly causing importance error.

transformers/src/transformers/masking_utils.py

Line 585 in a5a0c7b

attention_mask: Optional[Union[torch.Tensor, BlockMask]],

Therefore, this PR is committed for solving above two problems. By adding flex-attention not supported on Ascend NPU logic in is_torch_flex_attn_available, and updating BlockMask type annotations to string format.

Fixes # (issue)
#38362

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?