1. Instead of calling `pip_main` as a function, we're using `subprocess`
   to call `python -m pip`.  PIP behaves differently based on the
   version of Python being run and so we want to run PIP with Python 2
   when we're in Python 2 mode, and vice versa.
2. We're exposing the full `requirements` dict instead of just the keys.
   This is to support dynamically changing between the Python 2 and
   Python 3 version of a library, e.g.:

   ```
   native.alias(
       name = requirement_name,
       actual = select({
           "@bazel_tools//tools/python:PY2": requirement2(requirement_name),
           "@bazel_tools//tools/python:PY3": requirement3(requirement_name),
       }),
   )
   ```

Use METADATA, not metadata.json

Temporary changes to support running PY2 and PY3 in parallel

That refactor was entirely unnecessary, rules_python *does* apparently
call the .par files with the appropriate python interpreter so things
work in Python 2 and Python 3 as expected.

Note that this *does* keep the change to expose `requirements` publicly.

Revert the earlier switch to subprocess.call