[Feature] Improve handling & documentation of encoders and categorical static covariates in .fit_from_dataset()

Describe the issue linked to the documentation I couldn't find a mention of this in docs but it appears that TorchForecastingModel doesn't use encoders when using fit/predict_from_dataset - it doesn't warn about it either during the call. This seems like a potentially big blind side for the user since it is a silent behavior. Since this method is pretty well documented otherwise, this seems like an unexpected caveat.

I wonder if this should be documented OR actually move encoder logic to be included in the _from_dataset method calls.

Sidenote: This leads to issues with TFTModel since it just skips training and errors on predict if datasets don't contain future_covariates.

Sidenote2: categorical feature embeddings with TFTModel are also not working as expected (essentially ignored AFAIK) when using the _from_dataset methods since the code uses feature integer index instead of names for lookups but this seems like a completely separate issue.