feat: activate Focal Loss, add Dice Loss + span-width weighting, OpenVINO INT8 pipeline

[ALI-AL-MARJANI]

Summary

1. Loss functions (gliner/modeling/loss_functions.py, base.py, trainer.py)

focal_loss_with_logits already exists in the codebase but is disabled by default
(alpha=-1, gamma=0). This PR:

• Exposes loss_type, focal_loss_alpha, focal_loss_gamma in TrainingArguments so
  users can activate it
• Adds span_dice_loss() — span-level Dice Loss adapted from Li et al. (ACL 2020),
  applied element-wise over the (B, L×K, T) logit tensor with ignore_index masking
• Adds use_span_width_weight flag: positive spans of width k receive w(k) = 1 + log(k+1) — zero inference overhead

Motivation: WNUT-17 has 187× more negative spans than positive entities (0.53%
positive ratio). BCE's gradient is dominated by trivial negatives. Focal α=0.25 delivers
+0.99 pp WNUT-17 F1; Dice delivers +0.70 pp.

2. Bug fixes (gliner/utils.py, gliner/modeling/encoder.py, gliner/model.py,

gliner/onnx/model.py)

• is_module_available() changed from __import__() to importlib.util.find_spec() —
  prevents optional packages (peft, tensorflow) from being eagerly imported, which caused
  OpenMP deadlocks on macOS ARM
• encoder.py: kwargs.pop("token_lengths", None) prevents crash on bi-encoder models
  that pass this GLiNER-internal kwarg to HuggingFace forward methods
• Lazy import of Trainer/TrainingArguments in model.py — avoids importing
  torch.distributed at module load time

3. OpenVINO INT8 pipeline (scripts/convert_to_openvino.py)

New script: ONNX → OpenVINO IR → INT8 weight compression via
nncf.compress_weights(INT8_ASYM).

• 2.35× CPU speedup, 4× model size reduction, no accuracy degradation
• Uses weight-only compression (not activation quantization) because GLiNER's ONNX graph
  contains If nodes with dynamic rank that the CPU plugin rejects during calibration-based
  quantization

Benchmark

Config	WNUT-17 F1	Δ
BCE baseline	50.09%	—
Focal α=0.25, γ=2	51.08%	+0.99 pp
Dice Loss	50.79%	+0.70 pp

Backend	Latency	Speedup	Size
PyTorch FP32	59ms	1×	721MB
OpenVINO INT8	25ms	2.35×	181MB

Model: knowledgator/gliner-bi-small-v1.0, 200 fine-tuning steps on CoNLL-2003, eval on
WNUT-17.