Region recognition support

[ankandrew]

New CCT v2 models with region recognition (supporting +65 countries). There are also bunch of other enhancements and fixes in this PR.

Model changes

• Added empty plates (from backgrounds, noise, textures, etc.) to recognize all padded plates (empty plate).
• Region has Unknown region trained mainly with synth data.
• Changed gelu to silu due to export issues with some lib versions.
• The v2 pre-trained models have seen almost x3 more data than previous v1 models.
• Added new (and more correct) attention_layout, (that way each head doesn't use the full projection_dim)
• In both pre-trained v2 models, the xs and s versions, have over 0.99 for val_region_macro_f1 on validation held split (containing more than 114_000 samples).