Remove subsample default in OrthogonalProcrustesAlignment, improve tests

[stes]

Multiple tests failed internally and on this repo (e.g., in #20 , #25, #26) due to a numerical issue in OrthogonalProcrustesAlignment when a too small value for subsample is used.

This PR fixes this by setting the default to None, which means that all points available in the embedding are used for alignment. This is the numerically best option. In other cases, warnings are emitted, when e.g. the values of subsample is smaller than the number of points in the dataset, or when the number of subsample is smaller than 1000 (a reasonable default).

This PR also improves the tests, and removes the atol value of several tests, or at least decreases it slightly to avoid false negatives in the future.

---

Fix https://github.com/AdaptiveMotorControlLab/CEBRA-dev/pull/647
Fix https://github.com/AdaptiveMotorControlLab/CEBRA-dev/pull/645
Fix https://github.com/AdaptiveMotorControlLab/CEBRA-dev/issues/644

[github-actions]

Docstring Coverage Report

Name	Total	Miss	Cover	Cover%
main.py	5	1	4	80%
config.py	6	4	2	33%
data/datatypes.py	5	2	3	60%
data/helper.py	6	1	5	83%
data/multi_session.py	12	5	7	58%
datasets/demo.py	8	3	5	62%
datasets/gaussian_mixture.py	3	2	1	33%
datasets/hippocampus.py	7	1	6	86%
datasets/save_dataset.py	4	1	3	75%
datasets/allen/single_session_ca.py	9	3	6	67%
distributions/discrete.py	9	2	7	78%
distributions/multisession.py	5	1	4	80%
integrations/sklearn/helpers.py	2	1	1	50%
integrations/sklearn/metrics.py	3	1	2	67%
integrations/sklearn/utils.py	6	1	5	83%
models/projector.py	12	9	3	25%
------------------------------------------	--------	-------	--------	---------
TOTAL	396	38	358	90.4%
(45 of 61 files omitted due to complete coverage)

RESULT: PASSED (minimum: 90.0%, actual: 90.4%)