resolved packing params not propagating; added support for other backbones

[vivekkalyanarangan30]

Summary

• Ensure span and token forward paths keep packing_config/pair_attention_mask alive all the way to the encoder.
• Overhaul Transformer.forward so packed block-diagonal masks are translated into what each backbone expects (generic additive mask, plus dedicated paths for DeBERTa, ModernBERT, and T5/MT5).
• Extend the packing code path (including the packed shortcut) to pass both standard and pairwise masks, and clean up T5 embedding/attention handling so packed inference works for MT5-based checkpoints.

Testing

PYTHONPATH=. python bench/bench_gliner_e2e.py \
       --model knowledgator/gliner-pii-edge-v1.0

PYTHONPATH=. python bench/bench_gliner_e2e.py \
       --model knowledgator/gliner-pii-small-v1.0

PYTHONPATH=. python bench/bench_gliner_e2e.py \
       --model knowledgator/gliner-x-small

PYTHONPATH=. python bench/bench_gliner_e2e.py \
       --model urchade/gliner_small-v2.1

PYTHONPATH=. python bench/bench_gliner_e2e.py \
       --model urchade/gliner_medium-v2.1

PYTHONPATH=. python bench/bench_gliner_e2e.py \
       --model knowledgator/gliner-bi-edge-v2.0

https://www.diffchecker.com/2kR276Om/
All differences in output is because of 6th place of decimal precision

Revised benchmarks

Model	Without packing (s)	With packing (s)	Δ time saved (s)	Speedup
knowledgator/gliner-pii-edge-v1.0	3.268	1.746	1.522	1.87×
knowledgator/gliner-pii-small-v1.0	6.392	2.682	3.710	2.38×
knowledgator/gliner-x-small	10.454	5.135	5.319	2.04×
urchade/gliner_small-v2.1	3.795	1.288	2.507	2.95×
urchade/gliner_medium-v2.1	6.891	1.626	5.265	4.24×
knowledgator/gliner-bi-edge-v2.0	3.039	1.183	1.856	2.57×

[vivekkalyanarangan30]

@Ingvarstep couple of steps were missing for propagating the packing config from Gliner through base to encoder, fixed it
Checked benchmarks, now the gains are visible. Added support for mt5 backbone as well.

[Ingvarstep]

Awesome, thank you @vivekkalyanarangan30 once again for your contributions.