Is your feature request related to a current problem? Please describe.
There is a growing trend in using large, pre-trained neural models for zero shot forecasting. There is a whole family of models coming out, some of them Open Source. It would be interesting to have a single framework (Darts) that one can use to compare the "custom trained" and the "pre-trained" models, especially because the data preprocessing capabilities of Darts are very useful.

Describe proposed solution
"Implement" any/all of the following models:
MOIRAI
Lag-Llama
Chronos
MOMENT

Describe potential alternatives
See above

Additional context
Since the general timeseries libraries like Nixtla are committed to closed source / single foundational models, Darts could become a go-to for OpenSource solutions.