Miautawn-auto-validate-by-history

This release allows for different constraint selection strategies after exausting constraint recall increments. It's realised the `fpr_budget_fill_strategy` parameter within `AVH` constructor. Available options:
* `max_recall` - Fills the FPR cap using the constraints that offer max recall.
* `min_fpr` - Fills the FPR cap using the constraints that offer lowest FPR.
* `balanced` - Fills the FPR cap using the constraints that maximise their original recall divided by their expected FPR.

This release includes the functionality of stationarity testing!

The new `time_differencing` parameter is used when defining the `AVH` object:
python
from avh.auto_validate_by_history import AVH

avh = AVH(time_differencing=0)


It can take the following values:
* `0` - do not apply any time differencing (default).
* `>0` - will apply time differencing using this as param lag.
* `"auto"` - will try to find the stationarity with different lag options by using `adfuller` tests.

This patch release optimises `KsDist` metric, thus the setups that use it, should run faster.
Tested on 10k sized datasets, PS generation for one column dropped from 3s -> 1s.

This is the initial launch of the package!
The package is functional, but is missing some features described in the original paper, namely:
* Stationarity testing:
* Current iteration algorithm assumes that the data is stationary
* Categorical data support:
* Current code has a lacking support for categorical data in terms of data quality transformations and metrics. Thus it's best used with numerical data.