Consolidate per-backbone demos into unified from_ldm / from_clean

Replace the twelve per-backbone demo scripts (from_clean_*, from_ldm_*)
and their shared _demo_common / _demo_from_clean_common helpers with two
single dispatchers that select the backbone via --backbone:

  - from_ldm.py    : text/class -> latent diffusion -> PiD pixel-decode
  - from_clean.py  : image -> VAE encode -> optional noise -> PiD decode

Extract the shared runtime into focused modules: cli_utils.py (parser
construction), decoder.py (PiD decoder load/run + VAE round-trip),
step_capture.py (diffusers step-end callbacks for capturing xt/x0).
Slim inference_utils.py down to the helpers the dispatchers actually
use and point rae_generation / scale_rae_generation at the new modules.
Drop the unused create_dataset.py and refresh verify_env's module list.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>