Intronic

- Changes default scoring behavior to include all (5', BPS and 3') regions, instead of the previous default of just 5' and BPS. The 3' region typically contains less differentiation between U2- and U12-type introns, but may help reduce FP and FN classifier calls in edge cases. Of course, it's also possible that it could also _introduce_ FPs and/or FNs, although in my experience using all three seems to be more conservative than not.
- Misc. minor internal changes.

- Fix `GridSearchCV` regression with newer versions of `scikit-learn` (>v0.22) (see issue 1)
- Due to `scikit-learn`'s inversion of a default flag in `GridSearchCV`, `intronIC` must now require `scikit-learn` to be at least `v0.22`
- This fix **breaks compatibility** with `scikit-learn` versions <`v0.22`

- Replace parent-child hierarchical clustering of annotation features with simpler, directed graph-based approach
- Fix occasional issues where parent genes of CDS/exon features weren't correctly identified

A number of changes to the underlying data in this release - the default PWMs have been changed to a slightly less-stringent set, which should leave most results relatively unchanged and deals with some edge-cases where the original PWMs were overly penalizing for certain base positions due to being built from low-N samples. Other changes include:

- Default 3'SS region shortened to [-6, 4]
- By default, the human U2-type BPS PWM is used instead of the on-the-fly version. A per-run PWM can be generated using `--generate_u2_bps_pwm`
- z-scores in the output have been adjusted to correspond to the entire dataset (previously, they were based on the training set only)
- Non-canonical introns by default now use whatever PWM is closest to their terminal dinucleotides if one is obvious (e.g. for `AT-TC` introns, this would be the `AT-AC` PWM; for `AT-AG` introns, `GT-AG` and `AT-AC` are equally close in terms of edit distance). Otherwise, the terminal dinucleotides will be ignored and the best PWM will be selected based on the geometric mean of the component scores from each PWM. This can be reverted to the old behavior using `--no_ignore_nc_dnts`

- Uses human U2-type BPS PWM (data from Pineda 2018) by default.
To restore the previous paradigm wherein U2-type BPS PWMs are
generated on-the-fly using the best match to U12-type BPS motifs
in likely U2-type introns, pass --generate_u2_bps_pwm.

- Add best U2-type BPS to `meta.iic` output file.
Previously, only the best U12-type BPS sequence was reported. In certain cases, it may be useful to know which U2-type sequence was used in determining the BPS log-ratio score.
- Reduce formatting stringency for custom PWMs
This should reduce headaches if folks are adding their own PWMs by ignoring case, etc.
- Add clause to terminate multiprocessing pool processes on forced exit
There were cases I'd noticed in my own usage when force-exiting (e.g. via `ctrl-c`) where zombie processes would persist. Wrapping the whole thing in a try/except/finally seems to eliminate the issue (limited testing).