[torch.compile] support all attention backends

[torch.compile] support all attention backends #675