add use_kernel_func_from_hub to enable function level kernel mapping

[kaixuanliu]

No description provided.

[danieldk]

Adding function support is really a great idea and nice work on the PR! Some changes I think would be useful:

• Have the decorator on the function and not the layer that uses it, since we want to declare that the function can be extended and this also allows re-use.
• Related to the previous point: avoid having to match the name of the function in the mapping to the member variable. The member variable name is an implementation detail, whereas kernel mapping name is a global. E.g., a layer might call the activation function self.activation, but doing @use_kernel_func_from_hub("activation") would pollute the global namespace with a non-descript activation (which might conflict if multiple modules use the same name), whereas @use_kernel_func_from_hub("silu_and_mul") is not ambiguous.
• Functions handled by the same code path as modules (not requiring another iteration over the modules).
• Being able to use Hub layers for functions (if the signatures match modulo self).
• Being able to use Hub functions for layers (if the signatures match modulo self).

I have made a proposal that I think should work for your transformers PR, but also implements the properties above: #188

[kaixuanliu]

@danieldk really creative idea for your proposal, great work! Will close this PR once 188 is merged. Thx again!

[danieldk]

Merged! Let me know if you run into any issues using this in your transformers PR.