Automunge

- ok last rollout helped me recognize kind of a flaw
- which was the requirement for reserved strings in normalization_dict
- for use to derive a normkey in a few postprocess functions
- in cases where don't know returned columns or might return an empty set
- so revisited the columnkey parameter passed to postproces functions
- and repurposed from use as a string header returned from upstream transforms
- to now constituting a list of headers returned from all transforms with same recorded category applied to an input column
- (that is same recorded category as returned in the process function column_dict)
- this greatly simplifies the derivation of a normkey in those postprocess functions where needed
- and as a side benefit eliminates the need for reserved strings in keys of normalization_dict
- thus completely eliminating a channel for error in user defined transformation functions
- to implement had to make a few tweaks to processfamily functions as well as a new required process_dict entry
- new process_dict entry 'recorded_category' details the category recorded by a transformation function in column_dict
- which may be different than the associated category populated in family tree
- since the same transformation function may be assigned to multiple process_dict category entries
- also a new postproces_dict data structure as columnkey_dict
- which is populated for each applied transformation as
- {inputcolumn : {recorded_category : categorylist_aggregate}}
- where categorylist_aggregate is assembled as a list of all headers returned from transforms with the same recorded_category to the same inputcolumn
- (as may constitue columns returned from redundant applications like in cases where same trasnform is redundaantly applied to same input column but with different passed parameters)
- this categorylist_aggregate is then passed to the postprocess functions as the columnkey
- which can be used to derive a normkey without need to validate required unique keys of normalization_dict
- the tweaks to the processfamily functions were twofold
- in automunge processfamily functions we now populate entries to postprocess_dict['columnkey_dict'] with each column_dict_list returned from processing function application
- making use of the new support function _populate_columnkey_dict
- and then in postmunge postprocessfamily functions prior to calling the postprocess functions access the categorylist_aggregate to pass as a columnkey from the columnkey_dict
- using the inputcolumn to be passed to postprocess function and the recorded_category derived from the category populated in family tree's process_dict entry
- updated the 13 postprocess functions in library that previously used the required unique normalization_dict entries to derive a normkey
- to the new simplified convention available by use of the categorylist_aggregate passed to the columnkey parameter
- also updates to support function that validates the process_dict to account for new recorded_category entry
- and support function to apply functionpointer entries to user passed processdict to account for new recorded_category entry

- ok trying to get ahead of a new edge case for tlbn transform from 6.11
- for cases when transform is being performed multiple times in the same family tree
- such as might be applied redundantly with different bincounts or different distribution segments
- previously we inspected the bincount in postprocess function to confirm the correct normkey
- where normkey is the column header with suffix used to access normalization_dict entries
- however now we have optional alternate buckets parameter that may take precendence over bincount
- so went ahead and added buckets to the normkey derivation, meaning now postprocess fucntion inspects both
- which resolves the edge case
- also added both of these parameters to the _check_normalization_dict validation function
- which is a support function that confirms any required unique normalization_dict entry identifiers are so
- (we have a handful of processing functions that make use of similar methods to derive a normkey)
- also added a few code comments here and there for clarity

- added a new preset to AutoGluon ML infill option
- to use 'optimize_for_deployment' which reduces required disk space
- applied by default unless user activates best_quality
- appropriate since user doesn't need auxiliary functionality, models are just used for inference

- quick fix
- found a bug in both DPmm and DPrt that was introduced in 6.6
- all better

- fixed a feature importance bug introduced in 6.9
- associated with removal of some deprecated parameters
- reverted tlbn mlinfilltype from exclude to concurrent_nmbr
- which is needed for use in feature selection
- added some comments to predictinfill of potential extension for tlbn
- fixed a tlbn postmunge bug introduced in 6.11
- associated with accessing normalization_dict in postprocess function
- also added inversion support for tlbn

- added some more detail to function description of _postprocess_textsupport
- regarding format of textcolumns entries needed for operation
- (this convention was source of bug fixed in 6.10)
- new parameter supported for tlbn transform
- parameter 'buckets' can pass as list of bucket boundaries
- otherwise defaults to False
- leave out -/+ inf for first and last bins those will be added
- buckets is an alternative to bincount for number of equal population bins
- buckets, when specified, takes precendence over bincount
- allowing user to perform influence evaluation on custom segements of feature set distribution
- also changed mlinfilltype for tlbn from concurrent_nmbr to exclude
- after realizing that ML infill was messing up -1 out of range convention
- much cleaner this way, for tlbn missing data is just grouped with out of range bucket for entire set