text-to-pose: Use more datasets

[AmitMY]

The sign-language-datasets package includes datasets that could be used for this task:

• dicta sign (dictionary - hamnosys / pose)
• sign2mint (dictionary - signwriting / videos. poses can be easily extracted)
• signtyp (signwriting / videos) - quite noisy, I don't think it is worth using.
• swojs_glossario (signwiriting / videos) - very high quality, but long sentences rather than single words. I don't think it is worth using.

There are also more datasets, not currently available in sign-language-datasets, that would probably be good to add:

• dgs lexicon (dictionary - hamnosys / videos) - The DGS Corpus includes a lexicon for all the glosses. When clicking on a gloss (skip all $s for ease) then we have this for example which includes hamnosys and a video. Sometimes, only hamnosys exists, but with references in the DGS corpus when this hamnosys was performed. This is very powerful to get multiple almost-aligned examples for the same sequence (could probably extract 50-200k+ samples), in a non-dictionary manner
• The polish sign language corpus includes many dictionary videos (example) with hamnosys