[WIP][Add] Mask2Former

[ariG23498]

Fix #1044

[ariG23498]

Hi @qubvel some questions before I start contributing to the core part of the model:

1. This model seems to have three parts, a pixel encoder, a pixel decoder and a transformer decoder. While I understand that I do not have to write the pixel encoder, as that can be directly retrieved from timm, the only bits I would need to contribute right now would be the two decoders.
2. Do we want this to be inference only at first? To make this happen my workflow would be to copy code from the transformers implementation and make the weight conversion (only if required) and pass an image to the model to see the semgentation maps correctly.
3. I am unsure about the processor. Do you think we should concentrate on that as well?

Let me know what you think.

[qubvel]

Hey @ariG23498, thanks for the questions, I tried to answer them, but let me know if anything is unclear

1. Yes
2. AFAIU the semantic segmentation model should be trainable with existing tutorials, otherwise, we can make a new tutorial (if there are any nuances)
3. We can use Albumentations for preprocessing - similar to what I used for segformer + to create a notebook on how to make an inference. See also the snippet for Segformer https://huggingface.co/smp-hub/segformer-b3-512x512-ade-160k