Allow using custom SDPA for non-float32 dtypes in llama demo (#5548)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/5548

Converting the input to and from float32 is faster than not using the op. h/t to torchchat, which does this already (though it had a bug, which I sent a patch for).

Reviewed By: kimishpatel

Differential Revision: D63158951

fbshipit-source-id: 58c90d141ee403536c03a3b731f8547790fc9440