Implement and stress-test a byte-level BPE tokenizer

The tokenizer is one file, src/bpe/tokenizer.py, with three functions on top of a small Tokenizer dataclass that holds the merges in the order they were learned. Everything else is measurement: a script that downloads a public-domain novel, one that trains three vocabulary sizes, and one that counts tokens against o200k_base. The decision that matters most is keeping rank (when a merge was learned) as the only thing encode looks at: frequency belongs to training, never to encoding.