flash feature refactor

[michaelfeil]

This PR extends the Qwen2 architecture to other models other than Alibaba-NLP/gte-Qwen2-7B-instruct, given that the prior implementation was only covering such cases so as to use causal attention on CUDA and to rely on the provided tokenizer rather than patching it.

What does this PR do?

• makes flash-attn-3 and flash-attn-cpu easier to add.

Fixes # (issue)

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?
• 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.
• Did you write any new necessary tests? If applicable, did you include or update the insta snapshots?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[kozistr]

looks good to me!

[kozistr]

It's a nit; how about moving line 121 right after line 126 to the early-return non-cuda case?

Suggested change

	let flash_attn_enabled = &std::env::var("USE_FLASH_ATTENTION").unwrap_or("true".to_string()).to_lowercase() == "true";
	if cfg!(not(feature = "cuda")) {
	// if not cuda support, always false for now.
	return false;
	};
	if cfg!(not(feature = "cuda")) {
	// if not cuda support, always false for now.
	return false;
	};
	let flash_attn_enabled = &std::env::var("USE_FLASH_ATTENTION").unwrap_or("true".to_string()).to_lowercase() == "true";