Reduce the TQDM related output of NeuralForecast models

[SinghJivjot]

Description

I would love to help contribute to this NeuralForecast. I have been using this extensively for the last 2 months. I had a few things in mind of what we can add (inspired by AutoGluon)

1. A lot of unnecessary output is thrown in Jupyter Notebook (VSCode) while training Neural Models. Firstly reduce the junk TQDM output. Then put a option within the 'fit' method to toggle the display of progress bars. I don't want to use %%capture as it messes with the Pylance extension. I know this does not matter that as it will go eventually into Production, however while exploring the dynamics of NeuralForecast it would save time on scrolling. (I am sorry, I had kept this point for later but this takes precedence after I remembered it)
2. Adding {i+1} to epochs while training display instead of {i} :p
3. Creating a Stacked Ensemble of the Best Performing Neural Architectures

Use case

No response