Automunge

- two new categoric encoding options incorporated into library
- with transformation categories 'smth' and 'fsmh'
- these borrow from the label smoothing options previously available for label sets
- and allow them to be applied to training data categoric encodings in addition to labels
- accepts parameter 'activation' to designate the value for activations
- as float between 0.5-1, defaults to 0.9
- smth applies a one-hot encoding followed by label smoothing operation
- fsmh applies a on-hot encoding followed by a fitted label smoothing operation
- where fitted smoothing refers to fitting the null values to activation frequency in relation to current activation
- more info on label smoothing and fitted smoothing noted in essay "A New Kind of ML"
- (we still recomend the prior label smoothing parameters for target categoric labels in order to distinguish between smoothing as applied to train / test / validation sets)
- inversion supported with full recovery
- also found and fixed a small bug in fitted label smoothing

- inspired by the success of 5.90, a further simplificaiton to categoric defaults under automation
- now removed a kind of weird singluar scenario for training data sets with 3 unique entries which were treated with one-hot encoding
- and instead treated them to binarization consistent with other categoric sets
- also increased defaults for numbercategoryheuristic from 127 to 255
- (numbercategoryheuristic is the size of unique value counts beyond which sets are treated to hashing instead of binarization under automation)
- 255 unique values returns an 8 column binarized set (1 activation set is reserved for missing data)
- this update does not impact backward compatibility

- a big simplification to label set encodings under automation
- realized had accumulated too many scenarios, this way much clearer
- now quite simply, numeric data is given pass-through (no normalization), categoric data is given ordinal encoding (alphabetical sorted encodings)
- other label encoding options documented in new section in library of transformations in read me
- also small bug fix in feature selection originating from new convention of single column sets returned as series

- update to convention for returned sets
- now single column pandas sets are returned as series instead of dataframes
- this decision was based on conventions of some downstream libraries for receiving labels
- kind of like how numpy arrays need to be flattened with ravel
- also small tweak to NArw update from last rollout to reduce memory overhead

- a housekeeping cleanup to processing function naming conventions
- had included the suffic '\_class' dating back to very earliest experiments
- in hindsight this may have potential to be a point of confusion
- so scrubbed that suffix
- processing function naming now follows convention process_ / postprocess_ / inverseprocess_
- where is the transformation category returned in column_dict
- much cleaner this way
- also found a potential edge case channel for inconsistent processing between train and test associated with NArw aggregation for infill
- originating from NArow assessment overwriting entries for NArowtypes positivenumeric, nonzeronumeric, nonnegativenumeric
- cleaned that up, issue resolved

- Now when passing a processdict entry to overwrite an internally defined processdict entry, you can pass the functionpointer to point to itself, and then only have to populate the entries you are overwriting.