Automunge

- another oportunity found to reduce memory overhead
- this time associated with the getNArows function associated with recording infill entries
- such that now the evaluation is applied directly to a received column instead of a copy
- which has a small tradeoff in that projections to numeric types (such as force to numeric or force to positive numeric) applied in the function are carreid through the received column
- which means that for multi-transform sets categories with different NArowtypes will have a master NArowtype cast from root category
- this was deemed an acceptable tradeoff and is consistent with the assigninfill approach

- a slight reorg of some of MLinfill support functions
- which were originally built around scikit-learn which accepts numpy arrays as default
- as am adding autoML library options finding that some prefer pandas dataframes as input
- so moving the numpy conversion for scikit into the scikit specific support functions
- which in practice means also moving the ravel conversion and also reworking some of support functions from numpy to pandas basis, including some feature importance support functions
- weirdly finding that there is some kind of distinction between renaming pandas headers to integers vs conversion to numpy and back to pandas
- which currently comes up in autogluon support functions
- tabling that inquiry for now
- what's in place works for whatever that's worth
- also found and fixed an edge case for numeric extraction transforms addressed by additional conversion to numeric

- removed an edge case printout associated with direct pass-through columns
- printout was there to support troubleshooting
- no longer applicable after new inplace convention for excl transform
- added a filler entry for exc6 back into default transformdict and processdict
- to avoid troubleshooting printout activated with prior code demosntrtations which incldued exc6

- a revision to the NArw transform to reduce memory overhead (comparable functionality)
- added some printouts for remote edge case associated with inaccurate processdict lagelctgy entries to support troubleshooting

- some additional testing determined that the 5.15 update impacted the scenario when user externally defines custom transformation functions passed as entries to processdict
- so quick fix to accomodate this scenario
- (basically just means an additional type test for externally defined functions)
- also found and fixed an edge case for ML infill associated with inplace operations

- a few tweaks to the family processing functions to allow passing processdict entries without entries for (dualprocess / postprocess) or (singleprocess) for cases when not applicable
- so the new convention is only at least one of these two sets of entries for a processdict entry is required
- instead of both sets with "None" for those without use as was convention prior
- for cases where user specifies entries for both sets (dualprocess / postprocess) and (singleprocess), the (dualprocess / postprocess) takes precedence
- instead of checking != None now the entry is tested for validatity as a python function type which makes much more sense
- update to support function associated with functionpointer to accomodate these revisions
- also update to functionpointer support function to carry through additional designations to pointer recipient iff not previously specified, including inplace, NArowtype, MLinfilltype, and labelctgy
- thus new convention is that when defining a custom processdict entry with functionpointer, no additional entries are required if intent is to just carry through all designations from functionpointer entry
- similarly for chains of function pointers, the closest link for these items with entry takes precendence
- these updates all make things considerably simpler for a user, the goal after all