Avoid full vocab clone in get_vocab_size()

[eunseo9311]

get_vocab_size(true) was calling get_vocab(true).len(), which clones the entire model vocabulary into a new HashMap just to count entries.
For large vocabularies (e.g. LLaMA-3 128k tokens) this allocates ~10MB on every call.

Fix by computing base + added.len() - overlapping directly, where overlapping counts added tokens already present in the model via token_to_id. Zero allocation.

Applies the same fix to the Node binding, which had the identical pattern.

Adds a test covering the overlap scenario (token in both model vocab and added_vocabulary).

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.