[ENH] Make TimeSince trafo faster by changing period diff calculation

[KishManani]

What does this implement/fix? Explain your changes.

Inspired by how @ltsaprounis calculated the integer time difference when using a PeriodIndex in the FourierFeatures transformer, I've switched to using this alternative method which is much faster. I measured a factor of ~100 speed up in _transform.

An example to measure the difference:

from sktime.utils._testing.hierarchical import _make_hierarchical
from sktime.transformations.series.time_since import TimeSince

df =_make_hierarchical(hierarchy_levels=(10000,), min_timepoints=200, max_timepoints=200)

trafo = TimeSince(freq="D")
trafo._fit(df)
trafo._transform(df)

This PR will have conflicting changes with #4015. Once #4015 has merged I will resolve the conflicts in this PR.

What should a reviewer concentrate their feedback on?

Code quality. I'm having to write a lot of comments to explain the context of what's going on. I'm unsure whether this is a bad smell.

PR checklist

For all contributions

• I've added myself to the list of contributors.
• I've added unit tests and made sure they pass locally.
• The PR title starts with either [ENH], [MNT], [DOC], or [BUG] indicating whether the PR topic is related to enhancement, maintenance, documentation, or bug.

[fkiraly]

Merged #4015 so avoiding conflicts should now be easier.

[fkiraly]

I can confirm the speedup.
In a fresh python 3.10 env on windows, the example is reduced from 7 sec to 300 msec.

Functionality seems unaffected.

[KishManani]

It appears there are no conflicts 🤷‍♂️ , that saves some time then!

[fkiraly]

Re comments: I think this is a bit of a smell, but I would say, ultimately it's coming from pandas (and how it handles indices). The dependent code isn't bad.