Implement a minimal trainable byte-pair encoding tokenizer: learn merges, then encode and decode.

This question separates the field hard. The training loop is the core: repeatedly count adjacent pair frequencies across the corpus, merge the most frequent pair everywhere, record the merge, repeat. The tells are non-deterministic tie-breaking (two pairs tied in frequency must resolve the same way every run), applying merges out of learned order at encode time, and failing to round-trip decode. A candidate who handles an unknown character at encode time by falling back to the base character vocabulary shows they understand why BPE never has true OOV.