move CUDA flags out of --per_file_copts into the cu_library macro
Summary:
This has a few advantages:
* changes do not discard the Bazel analysis cache
* allows for per-target overrides
Test Plan: Verified with `bazel build --subcommands`.
Reviewers: seemethere
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/79414
Approved by: https://github.com/malfet
Author
Michael Andreas Dagitses
