[ENH] Adapt rocket classifiers to return numpy arrays instead of data frames

[TonyBagnall]

currently, all rocket transforms are converted into pandas before return in transform, despite all operations being performed on numpy. This PR switches all the rocket variants to return numpy arrays in transform. The classifiers that use these transforms then just convert it back to numpy. It is much more natural to just not use pandas at all. There are two stages

1. Stop the conversion in transform (remove pd.DataFrame(X_)
2. Stop the automatic conversion in the base class ( super(Rocket, self).init(_output_convert=False))

This PR is part of #208. Note I have left the unequal length version for now. For a bonus 10 pts, I will also

1. speed up the tests prior to merging. they currently take 2 mins locally for 6 rocket tests (even after numba build)
2. extend the docstring for rocket, which was a bit terse.

I have reverted the MultiRocket because the base class conversions are breaking it with 2D numpy. It breaks them all, but it makes MultiRocket fail tests. We can wait to do this until after we have resolved the conversion problem.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

change made