Add Tokenizer-Feature Logit Calibration (TFLC)

Novel eval-time mechanism: learns 6 scalar biases from byte-length
token classes (control/1byte/1byte+space/2byte/3byte/special) using
scored context, applies calibration to future logit predictions.

Nearest: ETLB (openai#1399, full vocab bias vector). Mechanistic difference:
TFLC uses 6 class-level biases from tokenizer metadata, not per-token.
Compliant: prefix-only, score-before-update, single pass.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: resouer