docs: clarify tokenizer considerations for new-language finetuning

[mukeshr06]

Summary

Addresses #225 by adding a focused "Finetuning on a new language" section to the existing fine-tuning guide.

• Explain why Falcon-7B's byte-level BPE can encode Chinese without adding vocabulary entries, and why encoding support does not imply language proficiency.
• Clarify why embedding parameter counts do not grow with new training text and how full finetuning differs from standard LoRA.
• Describe what intentional vocabulary expansion requires: stable token IDs, compatible embedding/output dimensions, training and saving the added-token rows, and matching tokenizer/checkpoint artifacts.

This is documentation-only. It does not introduce an automatic resizing API or change training behavior.

Validation

• pytest --check-links tutorials/finetune.md --check-links-ignore "http*": 5 passed.
• codespell tutorials/finetune.md: passed.
• git diff --check: passed.
• Loaded the published tiiuae/falcon-7b/tokenizer.json with Hugging Face tokenizers; verified exact encode/decode round-trips for Simplified Chinese, Traditional Chinese, and mixed English/Chinese samples, in-range token IDs, and an unchanged 65,024-token vocabulary.
• Cross-checked the embedding dimensions, LoRA freezing, and adapter-only checkpoint saving against the current implementation.

No model weights were downloaded and no training or full test suite was run.