docs(VisualBuilderPlan): F-G expanded to Universal Tokenizer Playground

Per user feedback: not just our non-standard tokenizer, but ANY
tokenizer with side-loading + multi-tokenizer side-by-side compare.

§9.5 rewritten with new structure:
  - 9.5.1 Tokenizer loader — 5 sources: preset library (12+ built-in:
    GPT-4/3.5/2, Claude, Llama-3, Mistral, Gemma, Qwen, DeepSeek-V3,
    Phi-3, cppmega_v3, nanochat_v3), HuggingFace Hub URL, local file
    upload (.json/.tiktoken/.model), paste JSON, custom tiktoken
    encoding (BPE ranks + specials + regex)
  - 9.5.2 Multi-format runtime registry with lazy-load:
    * HF tokenizer.json → @huggingface/tokenizers (8.3 kB, default)
    * tiktoken → tiktoken/lite WASM (~300-700 KB, lazy)
    * SentencePiece → @sctg/sentencepiece-js WASM (~600 KB, lazy)
    * Multi-provider → @goliapkg/tiktoken-wasm (~3 MB, opt-in)
    JupyterLite payload stays <1MB until WASM-heavy preset triggered
  - 9.5.3 Side-by-side A/B/C compare (up to 3 tokenizers):
    same input, per-row chips, cross-panel highlight bus (exBERT
    pattern — not in stock playgrounds), diff metrics dashboard
    (count delta / Venn / bytes-per-token / cost)
  - 9.5.4 Tokenizer-aware special-token classifier:
    FIM (red), CODE (blue), CONVERSATION (violet), DOC_SEP (green),
    MTP (amber), Engram (cyan), whitespace-sentinel (yellow lossy),
    generic special (grey fallback) — applied to ANY tokenizer with
    per-tokenizer overrides
  - 9.5.5 Greenfield layers (~800 LoC, up from ~600)
  - 9.5.6 New API endpoint with primary + compare_with[] + diff_metrics
  - 9.5.7 UI layout: A/B/C tokenizer pickers, three-column token chip
    panels, diff metrics row, doc-packing strip (primary only)

F-G ticket (cppmega-mlx-o0k.7) updated with same scope.

Budget: F-G 2 weeks (up from 1-1.5); total effort 10 weeks (up from 9.5);
wall-clock still ~8 weeks (F-G parallelises with F-C/F-D).

Author: apstenku123