feat: auto-pick GPU-appropriate precision config for full-precision models

When a full-precision model (diffusers base or all-in-one 全家桶 checkpoint)
loads with precision_config=[auto-derive] and no explicit config, identify the
model arch (reusing the arch+kind already fingerprinted onto xfm_ref) and load
the matching ModelScope series' precision config for the SELECTED GPU via the
precision auto loader — FP4 50x-above on Blackwell SM120+, INT4 50x-below
otherwise. Already-quantized inputs (nunchaku NVFP4/INT4, SVDQ, QuantFunc
exports) and explicit precision configs are left untouched.

- nodes_format_adapters.py: _ARCH_TO_SERIES, _FULL_PRECISION_KINDS allowlist,
  _device_sm (per-device SM with 0-safe fallback), _autopick_precision_for_full_model
- model_auto_loader.py: register Klein-4B/9B series (4B/9B share one precision-config)

Verified: real-scenario test across Klein 4B/9B, Z-Image, Qwen AIO + quantized
inputs (all pass); 4-dimension code review all-GO.

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