[inductor] Support select_algorithm with cpp_wrapper (#103003)
Summary: This is one step towards getting cpp_wrapper work with max_autotune.
Switch to use unique kernel name to cache generated cubin file.
This is a copy of https://github.com/pytorch/pytorch/pull/102738 to solve a ghstack issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103003
Approved by: https://github.com/jansel
