ENH: Add serialization to FramesStandardizer, remove saving with joblib

[NickleDave]

Using joblib to dump a fit FramesStandardizer leads to backwards-compatibility issues, e.g., because of a name change, as in #810

We should add read and write methods that just np.savez and np.load the arrays used to disk.

This doesn't magically solve errors that result from a name change, but it avoids the problem of needing the old environment around to unpickle the object. The numpy array format is super stable so we should rely on that instead.

Probably fine to just save it as a single-file, "standardizer.npz", don't bother doing something extra clever like saving in a directory.

[NickleDave]

I'm running into this issue when replicating previous results

Just making note that joblib definitely requires the original environment when unpickling, see table in sklearn docs here:
https://scikit-learn.org/dev/model_persistence.html#model-persistence

So this is a strong 👍 for making this change.

We should also:

• add migration guide to docs explaining how to handle this -- use snippet from Error using StandardizeSpect from training on previous version of vak #810 to explain how to migrate from files saved by earlier versions that used joblib