updates the three places in bucket reshard that write to bucket instance metadata: init_target_layout(), revert_target_layout(), and commit_reshard()

ECANCELED errors from cls_version are handled with a retry loop similar to retry_raced_bucket_write() from rgw_op.cc. each time we see ECANCELED, we re-read the bucket instance metadata and check for races with other reshards or reshard cancelation, then re-apply the mutation to bucket_info

i chose not to reuse retry_raced_bucket_write() here because it requires rgw::sal::Bucket, and reshard doesn't use sal interfaces yet - switching that would really complicate backports

qa/workunits/rgw/test_rgw_reshard.py has reshard test cases with fault injection to test error paths and crash recovery. i extended the test and radosgw-admin to allow injection of these ECANCELED errors. the injected ECANCELED errors are special because they're only injected once, so the test expects the retry to succeed