Adds optional scale and bias to cudnn's layernorm

[vedaanta]

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos and docs improvements)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure to update the docs?
• Did you write any new necessary tests?

What does this PR do?

This PR adds more functionality to cudnn's layernorm

1. Adds optional bias and scale
2. Updates to cudnn 9.1 requirements
3. Removes older way of caching full graph making function

Future:

1. Add benchmarks for LN fwd with litgpt configs
2. Add LN bwd
3. Add RMSnorm

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

[crcrpar]

when would we hit this? I'm not sure if raise is appropriate here at glance

[vedaanta]

If cudnn is not able to support the graph, this checker should return False. Other errors thrown will most definitely be user errors or gaps in cudnn support. Earlier cudnnex just used to return False for other errors too, but now I want to move it to raise them.

For example, right now a couple of thunder test cases do fail and it is due to a bug in latest cudnn. A workaround right now will be to block that failing layernorm config here in cudnnex explicitly. If the checker just returned False, it would have been hard to uncover this bug.

So these raise makes sure that both user code errors and bugs in cudnn get propagated to the user.
(@wujingyue suggested this workflow when working on cudnn's sdpa)

[lantiga]

tagging older draft PRs as later, feel free to reopen if this gets back being active