Add a targeted HFTokenizer encode-latency benchmark

[digantdesai]

Stack from ghstack (oldest at bottom):

• Make ReplaceNormalizer::normalize O(N) (single forward pass) #196
• Make HFWord::merge_all O(n log n) #195
• -> Add a targeted HFTokenizer encode-latency benchmark #194

Standalone Release-optimized benchmark for the HFTokenizer long-prompt encode
path. It repeats prose/code/dialogue templates to long inputs (~0.5k-8k tokens)
and prints mean encode latency. This is a targeted latency benchmark for this
code path, not a generic tokenizer harness; correctness is covered by the unit
tests and it exits nonzero if any encode errors.

Baseline latency on the original (pre-fix) code, Gemma-4-31B tokenizer, Release,
mean of 5 reps:

  vector          chars     ids     mean_ms
  prose_0.5k       2240     491       241.4
  code_1k          4514    1891       789.2
  dialogue_1.5k    6720    1345      2212.5
  prose_2k         8946    1450      3905.4
  prose_8k        49416    8005    123504.8

Latency grows with the square of input length (chars x5.5 from prose_2k to
prose_8k -> time x31.6 ~= 5.5^2); an 8k-token prompt takes ~123 s. Gemma's
normalizer turns spaces into the word marker before the space-splitter runs, so
the whole prompt is BPE-merged as a single piece.

Authored with assistance from Claude Code.