Support multilingual (mdeberta-v3-base) checkpoints

[julienmarie]

Summary

Makes multilingual GLiNER2 checkpoints work — e.g. fastino/gliner2-multi-v1 (mDeBERTa-v3-base backbone, fr/en/es/de/it/pt). Before this, such models loaded without error but decoded to garbage spans (fragment text, ~0.5 confidences); only the English deberta-v3-base base model produced correct output. This is the "Additional GLiNER models … not yet available" item from the README.

Diagnosis

The bug was isolated by comparing intermediates against the Python gliner2 reference on identical input. For fastino/gliner2-multi-v1, all of these were byte-identical to Python:

• the built input_ids (schema + text),
• the encoder hidden states (mean/std and per-token vectors),
• the subword→word pooled embeddings,
• the span_rep output.

That narrowed the divergence to the count-aware scoring head, where count_embed projects the label embeddings before the span-scoring einsum.

Changes (4 fixes)

1. vocab_size is read from encoder_config/config.json. The top-level config.json omits it, so it fell back to the default 128011 and the word-embedding was built too small for the multilingual 250112-token vocab. (GLiNER2.fromPretrained, ExtractorConfig.vocabSize made mutable.)
2. downloadModelDirectory now fetches encoder_config/config.json (added to the snapshot file patterns) so the value above is available.
3. The count layer is built per config.countingLayer. count_lstm → CountLSTM (GRU + MLP projector); count_lstm_v2 → CountLSTMv2 (GRU + DownscaledTransformer). It was hardcoded to CountLSTMv2, so count_lstm checkpoints loaded those (absent) transformer weights as random, silently garbling every span score. A small CountEmbedding protocol abstracts the two variants; a count_embed.projector key-remap entry is added.
4. safeSpans invalid-index zeros are Int32, not Float. MLX.where was promoting the span indices to Float, which made the downstream gather fail with "cannot gather with indices type" on the CPU backend.

Verification

fastino/gliner2-multi-v1 now matches the Python reference on both CPU and GPU:

EN  character: Cato 0.9996, Clove 0.9995, Thresh 0.9995, Peeta 0.992
EN  place:     District 12 0.97, cave 0.74, stream 0.67
FR  personnage: Charles 0.9999, Emma 0.999    lieu: Rouen 0.995

The English fastino/gliner2-base-v1 continues to produce identical results to before (the count-layer switch defaults to CountLSTMv2).

Notes

• Relation extraction for these checkpoints is still the separate WIP item.
• The README's "currently only deberta-v3-base is supported" line can be relaxed.