feat(loader): Klein 4B/9B original-precision auto-download + per-generation worker-log print

- model_auto_loader.py: add Flux.2 Klein 4B/9B to the original-precision base-model
  auto-download loader. A black-forest-labs/"klein" ModelScope search filter in
  _BASE_MODEL_SEARCH_CONFIGS + 4 explicit _BASE_MODEL_FALLBACK repos
  (FLUX.2-klein-{4B,base-4B,9B,base-9B}; "base" = non-distilled, no "base" = distilled).
  All 4 verified diffusers-format (model_index.json + scheduler/text_encoder/tokenizer/
  transformer), so the generic download_base_model_to_diffusers handles them as-is.
  The Klein 3-tier precision auto-pick (SM120 FP4 / SM89+ INT4+FP8 / else INT4+INT8 in
  _autopick_precision_for_full_model) already exists; 4B/9B share one precision-config
  (byte-identical, layer-name keyed) so the existing Flux2Klein->Klein-4B-Series map is correct.

- nodes.py: print "[QuantFunc] Worker log: <path>" on EVERY generation (WorkerManager
  text_to_image + image_to_image, after _ensure_worker) — the worker persists across
  generations so the one-time startup print scrolls away.

Multi-dim CR (correctness / regression / generality+simplicity): all GO.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>