fix(#449/#450/#446b): drop stray input_channels=4 from base QwenImage VAE config

bin/vae_configs/QwenImage.json carried a QwenImageLayered (RGBA 4-ch)
characteristic — "input_channels": 4 — on the base QwenImage 3-ch VAE. The SVDQ
adapter (nunchaku_svdq.py) UNCONDITIONALLY overrides the staged VAE config with
bundled_vae_config(arch) (it does not prefer a sibling config.json like
hf_native), so svdq QwenImage staging wrote input_channels=4 → engine
out_channels = value("out_channels", value("input_channels", 3)) built
conv_out=[4] → fail-loud vs the real [3] weights ("conv_out.bias dst [4] vs
src [3]"). The lighting/hf_native path prefers the sibling config (no
input_channels) and was unaffected.

- bin/vae_configs/QwenImage.json: remove the stray input_channels=4 (base 3-ch
  VAE → engine defaults out_channels=3; now mirrors QwenImageEdit.json)
- bin/vae_configs/QwenImageLayered.json: NEW, input_channels=4 — gate the 4-ch
  value to the genuinely-layered arch (correct bundled fallback for the RGBA VAE)
- format_adapters/factory.py: _STAGING_FORMAT_VERSION 3->4, invalidates any
  v3-staged dirs that baked the old input_channels=4

Verified on 远程-windows (SM120) via --engine comfyui qwenimage-svdq: VAE loads
(zero conv_out errors), coherent watercolor-cat image, status=passed.
run 20260625-svdqfix449-win.

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