generate.py: --save-intermediate / --load-intermediate for fast bake iteration

[NoahBPeterson]

Summary

Adds two flags that decouple the expensive sampling+decode phase from
the texture bake, so bake-only iterations cost ~1 minute instead of
~20+ minutes.

--save-intermediate PATH
    After pipeline.run() succeeds, dump the decoded mesh and voxel
    attributes to a torch .pt file. Same run still produces the GLB;
    this is a side effect that captures the state the bake needs.

--load-intermediate PATH
    Skip pipeline construction (the 90-200s cold load) and the full
    sampling+decode loop (~15-20+ min on M1 Pro). Load a previously
    saved blob and proceed straight to texture bake. The image arg
    and any --pipeline-type / --seed / --steps values are ignored —
    the cached mesh already represents whatever those values produced
    when the cache was made.

Why

The bake has three knobs worth iterating on (--texture-size,
alpha-mode heuristic, decimation target) but is gated on a 15-20 min
sampling phase. With a saved intermediate, bake-only iterations drop
from ~20 min to ~1 min on M1 Pro.

Use cases:

• Comparing texture resolutions (--texture-size 1024 vs 2048)
  on the same generated mesh.
• A/B testing the alpha-mode auto-detection heuristic (relevant after
  to_glb: percentile-based alphaMode auto-detection (default OPAQUE-friendly) pedronaugusto/trellis2-apple#1).
• Trying different o_voxel.postprocess.to_glb(decimation_target=...)
  values without resampling.
• Fast iteration on any future post-bake processing (e.g. our work
  finding the appropriate alphaMode defaults; this would have been
  a 1 hr loop instead of 12+ hrs without these flags).

Smoke test

On a saved 1.44M-vertex intermediate (M1 Pro 16 GB):

Loading intermediate from .../T_mac.intermediate.filled.pt...
Loaded intermediate in 0.1s
Mesh: 1,442,251 vertices, 3,014,498 triangles
Bake time: 21s
Saved: .../T_mac_pr5_smoke.glb

vs ~22 minutes for the equivalent end-to-end run.

Implementation

• --load-intermediate makes args.image effectively optional. We
  keep the argparse signature backward-compatible (image stays
  positional required) but gate the existence check on
  args.load_intermediate is None.
• The saved blob is intentionally minimal — just the post-decoder
  mesh + voxel attributes that to_glb consumes. No pipeline weights,
  no RNG state, no input image. Typical size: ~250 MB for a
  1.44M-vertex / 2.99M-face mesh + 6-channel voxel attrs at fp32.
• Layout (the attr_layout slice dict) is preserved as-is. On load
  the blob is wrapped in a SimpleNamespace so downstream code
  accessing mesh_out.vertices, mesh_out.attrs, etc., works
  unchanged.
• Existing watchdog-error handling and empty-mesh checks are preserved
  on the generation path; the load path bypasses them since by
  definition the cached mesh is non-empty.

API impact

No breakage: both flags default to None, so existing scripts run
unchanged.

Related

Pairs nicely with #11 (re-enable cumesh on M1)
and pedronaugusto/trellis2-apple#1 (alphaMode auto-detection) for
quick verification of bake-side improvements without paying sampling
cost on every iteration.