Fix llama.cpp chat token ID parity

[Qubitium]

Summary

• pass HF-rendered chat token IDs directly to llama.cpp
• avoid re-tokenizing rendered special-token text or duplicating BOS
• normalize BatchEncoding/tensor/list chat-template outputs

Validation

• 24 focused engine tests pass
• real Llama 3.2 tokenizer telemetry shows HF, Evalution, and llama.cpp chat token IDs have identical SHA-256
• official llama.cpp GGUF sanity generation returns the expected answer