Add support for unknown scripts in NLLB

The NMT engine does not have good support for scripts that NLLB was not originally trained on. All characters would be added as individual tokens to the vocabulary resulting in poor translation quality. silnlp provides better support for unknown scripts. Here is an example of the silnlp config to enable this support:

  tokenizer:
    src_vocab_size: 0
    trained_tokens: true
    trg_vocab_size: 500
    update_src: true
    update_trg: true

When silnlp is configured in this way, it will train a new subword tokenizer on the target data using the specified trg_vocab_size. The new tokens are added to the NLLB vocabulary. This results in better translation quality for unknown scripts, then the default settings. Similar functionality should be added to the NMT engine. The engine could automatically detect when to enable this feature or it could be enabled using build options.

Start by locating the NMT engine's tokenizer configuration and the code that builds or updates the NLLB vocabulary. Compare its behavior with the supplied silnlp configuration, especially trained_tokens, update_src, update_trg, and trg_vocab_size. Done means unknown scripts receive subword-tokenizer support, with a documented choice between automatic detection and build-option configuration.