Refactor merge_samples() to safely cleanup in case of errors

[brimoor]

Refactors the implementation of merge_samples() so that, if something goes wrong in the $merge aggregations, the temporary indexes and frame fields that are created in order to facilitate the merge will be automatically cleaned up.

Prior to this change, if a $merge failed on a video dataset, the datasets involved would be in broken states until the following operations were manually performed on them:

# Cleanup after failed merge
dataset._frame_collection.drop_index("_merge_key_1_frame_number_1")
dataset._frame_collection.update_many({}, {"$unset": {"_merge_key": ""}})

The fatal error was the unexpected _merge_key values, but leaving extra indexes laying around is also a party foul.

Update: adds some additional try-finally blocks in various places throughout the codebase

[benjaminpkane]

LGTM. This is where transactions are valuable. So next time we have to consider database state and errors, I'd recommend we look into #1791 before we start adding more code like this.