Part of #396.

Adds custom vocabulary biasing for all speech models (Whisper, Moonshine).

This PR supersedes #436, which added the bias infrastructure to WhisperModel only. Based on the review, the implementation has been moved to the base Model class, so it generalises to all speech models.

Changes

• cactus/engine/engine.h: added vocab_bias_ field and set_vocab_bias()
  to base Model class so all speech models inherit it automatically
• cactus/engine/engine_model.cpp: merge vocab_bias_ into tool_constrainer
  bias map inside Model::decode() before passing to gb->sample()
• cactus/ffi/cactus_transcribe.cpp: parse custom_vocabulary and
  vocabulary_boost from options_json, tokenize each word, call set_vocab_bias()
• cactus/ffi/cactus_stream.cpp: same parsing for the streaming path
• tests/test_stt.cpp: added test_vocab_bias_base_class which verifies
  the full chain — JSON parsing → tokenization → bias map → decode

How it works

When custom_vocabulary is passed in options_json, the FFI layer tokenizes
each word and builds a token_id → boost map. Inside Model::decode() this
map is merged with the existing tool constrainer bias and passed to gb->sample().
Boost values are clamped to [0, 20] to prevent degenerate outputs.

Testing

All stt tests pass. Debug output from cactus_transcribe.cpp confirmed the full chain works:

[vocab_bias] parsed 3 words, boost=15
  word='Omeprazole' token_ids=46 76 595 424 4765 306
  word='HIPAA' token_ids=39 9139 5265
  word='Cactus' token_ids=34 34775