Add Florence 2 support

[LinasKo]

Description

This PR adds Florence 2 support for supervision.

Run it with transfomers and parse with from_lmm(sv.LMM.FLORENCE_2, result).

No extra documentation was added to from_lmm at this point.

Type of change

• [x] New feature (non-breaking change which adds functionality)

How has this change been tested, please provide a testcase or example of how you tested the change?

Usage and testing in Colab: https://colab.research.google.com/drive/19XoQZ6LRdTUysAHUp-3JsfT_67T0rffE#scrollTo=CwPzmI9wOSQ6

Any specific deployment considerations

Run the Colab in High-RAM mode.

Docs

• [ ] Docs updated? What were the changes:

[LinasKo]

I found there's some tasks I overlooked.

In addition to examples on their Hugging Face page, they have 5-6 more on their colab.

At the end of today I'm leaving this in a working-but-incomplete state: I've added all tasks I could find, updated the Colab, but am not yet happy yet.

• I know multiple places where the code can be simplified
• REGION_TO_CATEGORY and REGION_TO_DESCRIPTION boxes are drawn incorrectly.
• I haven't exhaustively tested the empty cases, or researched the possible outputs. In many cases, e.g. REFERRING_EXPRESSION_SEGMENTATION, I'm getting one large [[[polygon]]], regardless of what inputs I try.
• It's late, so I simply ran out of time

Planning to dig deeper, simplify and add unit tests tomorrow.

[LinasKo]

Hi @SkalskiP,

This now has:

• Support for segmentation
• Support for regions methods
• Unit tests

Where it's lacking is:

• While I looked into text prompt, I made no effort to check whether users can provide their class names into the transformers method. In a few cases, label lists of ["", "", "", ""] are returned. My implementation ignores those.
• In some cases such as REFERRING_EXPRESSION_SEGMENTATION, I always get [[[polygon]]], regardless of what image and prompt I try. I did figure out what the layers represent and left some comments.
• Instead of returning nothing, some methods return very narrow long polygons. We could do post-processing here.

I invite to look at the Colab and play around with the results.