reduce S3 uploader mem usage

[matush-v]

Originally, we'd recurse through the directory to find files to upload. Based on the heap dump, the depth could get large if many subdirectories existed and were checked. The biggest piece seemed to be the list of paths themselves.

The order we check the files shouldn't matter here; we just want to check all files within a dir. So this updates the logic so we get all file paths we need to check, put them in a list, and loop through each path to check the file. This removes the overhead of the recursive method calls and the copies of references to Set<Path> synchronizedToUpload and List<Path> toUpload.

However, it's possible that the bigger problem is in the number Paths we have in the toUpload list unrelated to how it's handled within the recursion. One solution to that would be to just upload in limited batch sizes (e.g. 100 items at a time). Let me know your thoughts

[matush-v]

I've tested various spreads of files and directory recursion and the benefits aren't as plentiful as expected. Good news is the changes are uploading properly.
I've given the code another read and realized we're passing in a set of all files we've uploaded during the run. The uploader only used the set to check if we've already uploaded a file so that it doesn't upload it again. Singularity essentially guarantees no uploader clash with the unique request name being in the file path. I've removed this piece since it's a lot of added overhead

[ssalinas]

Going to close this in favor of just the unused path set one. I'll be pulling that into another bug fix as well.