Fix token_lengths kwarg leaking into the labels encoder

[yuriihavrylko]

Fixes #370

Problem

Since 3a38d45 ("make gliner serving more efficient") the collator emits a token_lengths kwarg — precomputed CPU-side text lengths that avoid an attention_mask.sum().tolist() GPU→CPU sync. The uni-encoder text path pops it, but BiEncoder.encode_labels only removed packing_config and pair_attention_mask before forwarding kwargs to the labels backbone. The text-side token_lengths therefore reaches the labels encoder and crashes any plain HF backbone:

TypeError: BertModel.forward() got an unexpected keyword argument 'token_lengths'

This breaks every bi-encoder checkpoint (e.g. knowledgator/modern-gliner-bi-base-v1.0, knowledgator/modern-gliner-bi-large-v1.0) on transformers 4.x. Bisected to 3a38d45: every commit before it predicts correctly, every commit after crashes.

Fix

Pop token_lengths in encode_labels alongside the other text-side kwargs — the value describes text sequences and is meaningless for label sequences.

Verification

knowledgator/modern-gliner-bi-base-v1.0 + transformers 4.48.3:

• current main: TypeError above
• with this fix: [('Angela Merkel', 'person', 0.85), ('Kyiv', 'location', 0.75), ('CDU', 'organization', 0.93)] — identical to gliner 0.2.16 output

Note: on transformers 5.x this crash does not reproduce (the kwarg is silently swallowed), but bi-encoders fail there for an unrelated reason — see the companion PR for #324.