Switch all IO to fsspec

[guipenedo]

WIP, a better solution needs to be found for tokenization and other memmap dependent blocks for sources like s3.
Paths can be passed as simple strings, or as tuples (str, fs options dictionary) or as a DataFolder object directly.

addresses #41

[guipenedo]

Conflicts fixed

[guipenedo]

Should be ready now. Replaced most mmaps with simple seek on streams (there is quite a bit of jumping around, mostly on the shuffling part of tokenization). I tried to optimize block_size a bit (and disabled caching) for this part, but it's still much slower than before (write file to disk, memmap it, create new one by shuffling that previous one and only upload the final one)
Would love to hear your thoughts/comments @mariosasko

[mariosasko]

Nice! Regarding the mmap issue, I think we should be able to address it by using the LocalFileSystem's underlying file descriptor in the mmap functions.

[guipenedo]

Nice! Regarding the mmap issue, I think we should be able to address it by using the LocalFileSystem's underlying file descriptor in the mmap functions.

I will add a local "working_directory" for this part then I suppose

[mariosasko]

Two more comments