Automunge

- quality control audit:
- visual inspection of transformation outputs of entire library (passed)
- visual inspection of inversion operation of entire library (passed)
- visual inspection of complete data structure initializations for transformdict and processdict (see below)
- visual inspection of complete infill library operation (passed)
- quick walkthrough of various operations that inspect MLinfilltype (passed)
- reduced scope of import from collections library to just the counter module
- found and fixed incorrect processdict entry for spl2 labelctgy (this only comes into play when spl2 applied as label category during feature importance)
- removed spl3 and spl4 from library as spl3 was redundant with spl2 and spl4 was not very useful
- changed ors2 use of spl3 to spl2
- updated bxcx family tree to allow bins assembly downstream of nmbr if binstransform selected (equivalent to prior configuration without bins assembly available as bxc4)

- a few small tweaks to evalcategory function for efficency

- new validation to confirm feature selection model successfully trained, returned as check_FSmodel_result
- new validation to confirm passed numpy array is tabular (eg 1 or 2D), returned as check_np_shape_train_result, check_np_shape_test_result
- new one-hot encoding varient onh2, similar to onht but includes a NArw column by default even when NArw_marker parameter not activated
- (as one-hot by default otherwise returns rows without activation for infill)
- the thought, and just a hunch, but I expect there may be some libraries where labels may require all rows to have an activation, so yeah this is now available as an option for label encoding for this scenario
- some tweaks to onht/text/pwrs/pwr2 to remove use of a second (temporary) column suffix - one less opportunity for overlap error
- a little cleanup to postprocess_text and postprocess_onht
- add some clarification of model basis to READ ME for ML infill

- material update to transforms applying one-hot encoding and power of ten binning: text, onht, pwrs, pwr2
- this update breaks backward compatibility for these specific transforms
- corrected the order of returned columns to maintain consistency with order of received columns
- note that prior method still had consistent order between automunge and postmunge so passed validation
- but this way just makes more sense, especially to maintain order grouping when one-hot is part of a multi-output transform set

- found and fixed small bug for feature importance evaluation with binary classification
- originating from the 4.36 update

- added validations for column header overlap detection associated with Binary and PCA dimensionality reductions
- removed convention of different default categoric encoding when applying Binary transform
- (now that Binary transform can be applied on partial sets this made less sense)