Re-implement LM rescore for online transducer

[SilverSulfide]

Shallow fusion can be too slow for online cpu inference. Added an option to use classical LM rescore instead.

• Rescore implementation based on https://github.com/k2-fsa/sherpa-onnx/pull/133
• Shallow fusion enabled by default
• Pass lm-shallow-fusion=false to enable rescore instead
• Updated online Python API

CPU runtime comparison for ~1 min wav file using the default --lm-num-threads=1

Method	Runtime (s)	RTF
No LM	5.7	0.098
Rescore	40	0.68
Shallow Fusion	69	1.2