Automunge

- resolved an inversion printout associated with fragmented dataframe warning (just needed to move the defragmentation operation to a different spot)
- options to pass noise injection distribution parameters as scipy stats distributions now supports seeded sampling

- added a False scenario for numbercategoryheuristic parameter to deactivate the heuristic
- (numbercategoryheuristic sets threshold for number of unique entries in a feature beyond which hashing is applied)
- note that hashing still applied under automation in the all unique case
- new noise injection option as DPsk
- DPsk is a masking injection, that masks a sampled ratio of entries with a designated mask_value defaulting to the integer 0
- have seen mask noise discussed in multiple papers, wouldn't know who to cite, I believe this is fairly common practice
- note that the DPsk process_dict specification is structured similar to DPne and DPse as direct passthrough without data type conversion, infill, or NArw aggregation
- can also be integrated into a family tree for application in conjunction with other encoding options.
- updated convention rolled out in 7.74 for sampling noise parameters from list of candidates
- as previously implemented a noise setting was sampled in automunge and carried through to postmunge
- in revised convention a unique setting is sampled in both automunge and postmunge
- this update aligns with the prior convention for passing noise parameters as scipy stats distriutions

- fixed a typo that was interfering with seed populating in default sampling_type scenario
- corrected seed populating for transform_seed sampling_type scenario
- found and fixed a process flaw when populating seeds for sampling_seed sampling_type
- (basically identified that our seed populating was only working per specs with the bulk_seeds sampling_type, now all sampling_type scenarios working as intended)
- consolidated noise transform support functions get_nprandom and erase_seeds, previously were defined in each function, now only defined once (better practice)
- new option for noise transforms, assignparam specification of noise distribution parameters flip_prob/test_flip_prob/sigma/test_sigma now accepts parameter specification as a list of candidate values for automatic seeded sampling between candidates
- note that the automunge(.) established basis carries through to postmunge(.)
- found and fixed a bug with inversion list specification

- new assignparam option for numeric noise injection via DPnb as 'rescale_sigmas'
- rescale_sigmas defaults as False, when True the received sigma specifications are multiplied by standard deviation of feature as found in the training data
- note that DPnb default noise profiles were set to align with z-score normalized data which by definition have a standard deviaiton of 1
- the rescale_sigmas option allows user to defer to common default parameters even when injecting to non-normalized data
- DPne, which is for numeric noise injection to pass-through sets without feature normalization, now defaults to rescale_sigmas=True

- after attempting further validations on 7.71 don't feel comfortable putting autogluon option into production
- there are some subtle aspects of complexity that would need to consider before reintroducing
- making an executive decision to withdraw autogluon Ml infill option and also default infill option
- hoping this doesn't inconvenience anyone, based on 7.71 discovery that was running outdated imports expect this hasn't been a widely used feature

- identified that our autogluon ml infill option was based on a very old version
- which had applied a different import procedure
- updated autogluon to the current import procedure
- apparently our validations were being run on version 0.0.15, they are now up to 0.3.1
- autogluon should be working now
- pending formal validation, for convenience intend to validate after rollout