[attention backends] remove non-hub attention backends.

[sayakpaul]

What does this PR do?

Now that the Hub-based attention backends (Flash, Flash 3, SAGE) have been fully feature-complete (torch.compile, CP, etc.), we can confidently remove the non-Hub variants.

This has several advantages:

• More Hub-centric
• Eliminates complex installation nightmares
• Lesser maintainence burden (otherwise we would have two different code paths for each of the features for each attention backend)
• Helps us improve kernels as well

Cc: @danieldk @LysandreJik

[sayakpaul]

I will run the tests in test_attention_backends.py as well.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.