As discussed with @msaroufim, we plan to treat each overload (e.g., add(tensor, tensor) vs. add(tensor, scalar)) as a separate kernel.

Benefits:

• This approach makes it easier for both humans and LLMs to generate correct and efficient kernels, since they only need to handle one input case at a time.
• We expect this will increase engagement and improve the rate of correct kernel generation by LLMs.

TODOs:

• Split operator tests for each overload.
• Register kernels for specific overloads, rather than for all overloads at once.