Bugfix auto packing with streams + no remote path

[mattyding]

This PR

Edge case where auto packing fails for a streaming dataset specified through streams field if only a local path and no remote path is provided. This configuration is more broadly supported for Foundry.

  train_loader:
    name: finetuning
    dataset:
      streams:
        finetuning:
          local: /foo/bar/train/mds/

An error is thrown in StreamingDataset init because there is a check that local contains an index file. The current auto packing script overwrites local with an empty tempdir. This fix temporarily has remote point to the user-provided local, which should contain an index file.

Testing

• MCLI run with streams + no remote path
• Added unit test

[mattyding]

Re-tested that it trains E2E with the following cases:

• streams, local no remote
• no streams, local == remote
• no streams, remote only
• no streams, local only

[dakinggg]

CI failure looks unrelated, will have to figure that out separately