Support efficient attention

[daisyden]

🚀 The feature, motivation and pitch

Please support torch.nn.attention.SDPBackend.EFFICIENT_ATTENTION and torch.ops.aten._efficient_attention_forward.

Alternatives

main

Additional context

No response