BLEU: support tokenizer string alias and clarify input edge cases

[Vangmay]

This PR makes a few small, backward-compatible improvements to the BLEU metric implementation to improve usability and robustness.

Changes

• Support a string alias for the default tokenizer (e.g. tokenizer="13a") in addition to passing a callable.
• Add explicit validation for length mismatches and empty input lists.

These changes do not modify the default BLEU computation or scoring logic and are intended to make behavior clearer and easier to debug for users.

Fixes #729