Fix: Reinstate dateparser==1.2.2 with a workaround for the upstream issue

[erindru]

dateparser==1.2.2 introduced a regression that essentially causes the following to fail:

>>> import dateparser
>>> dateparser.parse("1 minute ago")
# (returns None instead of a datetime)

This breaks a bunch of tests in SQLMesh.

However, dateparser==1.2.2 also introduced a cache that shaves several seconds off SQLMesh startup time which can add up over many invocations of SQLMesh.

So this PR adds a workaround for the upstream regression until it can be resolved upstream. The workaround is simply deleting the dateparser_tz_cache.pkl pickle file that is distributed with dateparser==1.2.2 and allowing it to be regenerated on first load.

However, we need to be careful just to delete the one that was distributed from upstream and not one that has been generated locally, since it's the source of the startup time performance improvement.

[erindru]

This pin forces SQLMesh to use the version of dateparser that contains the performance enhancements (which dont exist in any other version), but also the parsing regression that this PR provides a workaround for

[erindru]

Closing in favour of keeping the <= 1.2.1 pin and accepting the performance penalty until dateparser is fixed upstream.