Change pull_paths_from_storage to use gitattributes style filters

[Panaetius]

If we have a workflow that creates thousands of files in an output directory, the gitattributes filter might look like data/raw-data/** filter=lfs diff=lfs merge=lfs -text.
When running a subsequent run with this directory as input, we call pull_paths_from_storage with ALL the files in the directory which in turn calls git lfs pull on them in batches of 100, which can take a pretty long time (1 hour for 100k files on my machine). But git lfs pull does accept filters, so instead just calling git lfs pull -I data/raw-data/** would only take 0.17 seconds.

For directory inputs, we should call pull with just the directory.

for testing, you can use https://ai.stanford.edu/~amaas/data/sentiment/ . Use renku run to extract the archive and then use something like cat extracted/aclImdb/train/* > output in a subsequent workflow.