[AOTI][cutlass backend] Do not remove the cutlass kernel .o file after packaging (pytorch#154155)

Differential Revision: [D75253009](https://our.internmc.facebook.com/intern/diff/D75253009/)

In general, we want to cache the cutlass kernels.

Also saw an error saying .o not found.

Pull Request resolved: pytorch#154155
Approved by: https://github.com/chenyang78

Author: henrylhtsang

torch/_inductor/codecache.py

@@ -2082,6 +2082,8 @@ def _pad_to_alignment(raw_bytes: bytes) -> bytes:
             else:
                 so_builder.build()
                 for o_file in obj_srcs:
+                    if o_file not in gpu_kernels_o:
+                        continue
                     # Remove these as they are not needed anymore
                     os.remove(o_file)