Coalesce related performance tests and correctness tests under a single op name

[PaliC]

Especially for test suites like op info, we have a single op_name (from

BackendBench/BackendBench/scripts/setup_operator_directories.py

Line 22 in adefb1d

def extract_operator_name(op_str: str) -> str:

) which corresponds to a bunch of overloads but more importantly single test sets.

As we do in llm-relay we should expect users to produce a single kernel per op_name, however, having different sets of tests for the same op name is a unintuitive (an example is div in opinfo). We should have our suites coalesce performance and correctness tests that correlate with the same op name.

[msaroufim]

Granted the sort of obvious consequence here is investigating if we should be making additions to op info and then doing more frequent and broader coverage operator and shape suites using torchbench infra