[FEAT] Implicit Quantile Networks

Features:

Adds Implicit Quantile Loss as a loss function to facilitate Implicite Quantile Networks for all architectures. This implementation is based on the GluonTS implementation, following the original Gouttes IQN paper.
Cleanup moving parts to base_model out of the individual architectures, removing exceptions out of individual architectures.
Adds a number of class attributes to models that indicate whether the model supports exogenous.
Fixes some architecture errors, mostly in *former models.

Tests:

References:

Nixtla / neuralforecast