STL decomposition

[paxcema]

Why

To improve the set of tools available for time series mixers to use if it is deemed useful.

How

This PR does a bunch of stuff, namely:

• Big refactor in the TS transform step. We now default to using pandas date/time utilities, meaning that indexes are now proper timestamps (multi-index in the case of grouped tasks). This both simplifies the logic and makes it more robust for things like inferring new rows according to the observed series frequencies.
• Introduces the NeuralTs mixer, which inherits from the vanilla Neural but adds time-series specific logic. This was done to keep specialized procedures separated from the (simpler) classification and regression tasks. JsonAI dispatch has been modified accordingly to make use of this.
• Available time series groups at training time are now logged and registered inside the StatisticalAnalysis object.
• Some simplifications to ConcatedEncodedDs, though we should probably remove this abstraction entirely (see Make EncodedDs interface simpler #746).
• Moves a bunch of TS utilities to lightwood.helpers.ts
• ts_analysis phase now finds (and fits) optimal deseasonalizer and detrender objects for each series available at train time, for which we use Optuna.
• The respective methods that mixers can use to leverage these STL blocks are in mixers.helpers.ts
• LightGBMArray, SkTime, Prophet have been updated to use STL blocks, controllable by JsonAI args if needed. (NOTE: NeuralTs calls the blocks but there's no effect because it uses encoded values inside the DS, this will be addressed in a new PR)
• Encoder TsNumericEncoder has been simplified to avoid storing the sign, instead simply applying the underlying normalizer. This seems to be especially useful for NeuralTs to achieve improved forecasts, but formal benchmarks are still pending.
• Modifications to the analysis.nc block to work with all the above.