Fix test failures with Pandas 1.2.0

[BryanCutler]

There are a number of test failures when using Pandas 1.2.0.

• The method pandas.core.ops._get_op_name() has been removed in Pandas >=1.2.0. This was being used when building TensorArray ops, and is replaced with similar usage in ExtensionScalarOpsMixin.
• TensorArray should return it's instance when astype() called with same type
• TensorArray._contains_ needs to have proper implementation of any()
• repr() is broken when using float values of ndims > 2

Fixes #158

[BryanCutler]

There seem to be a number of different issue with Pandas 1.2.0, I'll look into fixing them here too.

[frreiss]

@BryanCutler any idea what's causing that test failure on the Python 3.7 build (https://github.com/CODAIT/text-extensions-for-pandas/pull/157/checks?check_run_id=1620791768)?

[BryanCutler]

I fixed most of the issues with pandas 1.2.0 and TensorArray, except there is a regression with repr() with floats, similar to #151. There has been some back and forth with my fix upstream for that one, so maybe I can fix both. Unfortunately, I don't know of a workaround, so we either have to leave it as a known issue or cap the supported pandas version < 1.2.0. I'm looking into the other issues now.

[BryanCutler]

Not sure why this is different now, but I think it's safe to skip and I can make an issue to follow up with later

[frreiss]

Works for me.

[BryanCutler]

@frreiss this failure https://github.com/CODAIT/text-extensions-for-pandas/pull/157/checks?check_run_id=1646927381#step:5:1438 in bert.align_bert_tokens_to_corpus_tokens() seems a bit strange. Not sure why this would suddenly change, unless they added some optimization that wasn't complete. I can look into it further, but letting you know in case you have some idea what's going on.

[BryanCutler]

Ok, I think I found why the above error is happening and looks to be a bug. Pandas attempts to do a cython aggregation and then fallback to a vanilla agg if there is an error, but checks the error message improperly for this case. I should be able to file a bug report and do a PR for the fix.

[frreiss]

Thanks for tracking down the root cause of that problem, @BryanCutler !

[BryanCutler]

PR for the agg fix pandas-dev/pandas#38982

[BryanCutler]

@frreiss I capped the upper version of Pandas to < 1.2.0 because not being able to display a Series with TensorArray of floats is pretty major and I don't see a workaround. I'll keep working on the upstream fix, but what are your thoughts on merging this?

[frreiss]

I think merging is the best way forward for now. We may want to post a new release just to get that pandas<1.2.0 constraint into our requirements.txt.