Automunge

- A few improvements to simplify trainID_column and testID_column specification requirements
- For scenario when passing both train and test data to automunge(.)
- Now only have to specify trainID_column if ID columns are the same between train and test sets
- Changed conviention for postmunge(.) inspection of ID columns from postprocess_dict
- To inspect the train ID columns instead of test ID columns
- To be consistent with convention that postmunge(.) is based on properties of the automunge(.) train set

- added a data type retention check to postmunge ML infill to match automunge
- found a sort of redundant operation in sp10 that was interfering with drift stat population intent, removed the redundancy
- inversion now supported with partial recovery for splt / spl2 / spl3 / spl4 / spl5 / spl6 / spl7 / spl8 / spl9 / sp10 / sp11 / sp12 / sp13 / sp14 / sp15 / sp16 / txt3 / ors7 / ors5 / ors6 / ors2 / or11 / or12 / or13 / or14 / or15 / or16 / or17 / or18 / or19 / or20 / srch / src2 / src3 / src4 / nmrc / nmr2 / nmr3 / nmcm / nmc2 / nmc3
- decided to simplify the search options to just focus on srch and src4 so as to avoid distraction
- option to aggregate search terms into common activation rolled out for srch in 3.91 now available in 3.94
- new MLinfilltype 'concurrent_nmbr'
- intended for returned multi-column numerical sets
- such as may be returned from tlbn for instance
- previously these sets were excluded from MLinfill
- in this new type each column in the returned set has a distinct trained model
- updated MLinfilltype for tlbn to concurrent_nmbr
- new MLinfilltype 'concurrent_act'
- intended for multi-column boolean sets which may contain concurrent activations
- such as may be returned from sp15 for instance
- previously these sets were given comparable ML infill treatment to 1010
- such that a single model was trained for the collection of activations
- in this new type each column in the returned set has a distinct trained model
- updated MLinfilltype for sp15 ,sp16, srch to concurrent_act

- 3.94 rollout contained a silly mistake
- I had somewhat lazily overwritten the sp13 and sp14 and replaced them with new transforms
- and just renamed the original to sp15 and sp16
- this of course broke compatibility with prior family trees that had incorporated sp13 and sp14
- which I had forgotten about
- so this update reverts the original sp13 and sp14 to prior configuration and remade the new trasnforms as sp15 and sp16
- details matter

- removed a redundant intermediate floatprecision transform as rounding for 16 bit float scenario was interfering with validation
- now floatprecision transforms take place only at conclusion of operation
- realized the new splt/spl8 parameters for concurrent_activations neccesitates different MLinfilltype due to returned columns with multiple simultaneous activations
- so scratched the parameter option and incorporated that functionality into new transformation categories sp13/sp14
- also found an efficiency improvement opportunity for splt family transforms by incorporating a break in one of the for loops

- new 'concurrent_activations' parameter accepted for splt and spl8 string parsing transforms
- allows the returned boolean columns to have simultaneous activations for overlaps detected accross entries
- note that activating this parameter may result in high dimensionality of returned data

- found potential source of noise for encodings based on value counts
- relevant to ord3, bnry, bnr2, ucct, and lcinfill
- this was not impacting consistency between automunge and postmunge
- but it did have potential to demonstrate inconsistency between seperate automunge calls
- the issue was when grabbing value counts those entries with same number of counts did not have a sort method
- so now have incorporated a second basis of sort into value counts (first by count, then by alphabetical)