[Windows] Add hasattr checks for distributed to improve compatibility

[0xDELUXA]

Motivation

Without these checks, ROCm on Windows raises an AttributeError when importing certain distributed functions.

For example, when adding a custom node to ComfyUI:

  File "E:\ComfyUI\venv\Lib\site-packages\flash_attn\utils\distributed.py", line 12, in <module>
    torch.distributed.all_gather_into_tensor = torch.distributed._all_gather_base
                                               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: module 'torch.distributed' has no attribute '_all_gather_base'

Cannot import E:\ComfyUI\custom_nodes\RES4LYF module for custom nodes: module 'torch.distributed' has no attribute '_all_gather_base'

This PR adds hasattr checks to prevent this error. It is not expected to cause any issues on Linux.

Technical Details

Improves ROCm compatibility with FlashAttention-2 on Windows.

Specifications:
GPU: AMD Radeon RX 9060 XT
Python: 3.12.10
PyTorch: 2.11.0a0+rocm7.11.0a20260121 (TheRock)
Triton: 3.6.0a0.post25 (triton-windows)
OS: Windows 11

Test Plan

Functional testing on AMD Radeon RX 9060 XT (gfx1200).

Test Result

All existing Triton Flash Attention tests pass on gfx1200.

The AttributeError has been resolved.