[fbsync] Fixed file filtering bug in SST2 dataset

[Nayef211]

Summary:

• Removed copying partial SST2 asset file to a temp dir and instead directly working with the file from the asset folder
• Fixed bug with path names affecting how files were filtered out from the zip file
  • For example, if the value of split is "test", the following snippet of code filter(lambda x: split in x[0]) might match all of the "train", "test", and "dev" files depending on the location of the dataset asset file
  • When testing with buck, the location of the extracted files could look something like /data/users/nayef211/fbsource/fbcode/buck-out/dev/gen/pytorch/text/test/experimental_test_datasets#binary,link-tree/test/asset/SST2/SST-2.zip/train.tsv. Since the word "test" is contained in this path string, the filtering logic would incorrectly select the "train" file even though what we want is the "test" file
  • To resolve this we append the file extension (in this case ".tsv") to the split variable in the filtering logic

Reviewed By: parmeet

Differential Revision: D32329831

fbshipit-source-id: dbb4803a04f6cd50fab3f7ce5530d3258b2db012

[parmeet]

LGTM! Just put [fbsync] in front of the title, so that if we plan to import on per-commit basis, we know which commits to skip :). Thanks!