Some clean up for SharedWriter incidents for Metadata in multi-stream mode.

This changes reduces the number of incidents thrown by the SharedWriter when run in multi-stream mode (such as derivation). Rather than throwing input incidents for each output file from all the clients, it will collect all the metadata from a client for all files and then throw the incident. This should be very close to what happens in traditional AthenaMP & merge, except some unavoidable differences. This may help with Jira ATLASG-1477, but there seems to be no reproducer, so its hard to confirm. Tagging @jcatmore , @elmsheus : I did some basic testing, but would appreciate if someone else could verify we I am not breaking Derivation. Thanks!