feat: export sky and calibrated source-trajectory renders

[heiwang1997]

Summary

This addresses the sky-rendering request in issue #5's follow-up comment and makes the reference renderer honor the source camera calibration and trajectory.

• derive a renderable cubemap from the released decoder's canonical RGB and SKY semantics (the public checkpoints omit the learned sky.* head)
• retain the observation mask, infill unseen directions from pruned ROAD Gaussians, and merge chunk cubemaps with mask-aware weights
• export a versioned <stem>.sky.npz sidecar plus <stem>.sky.png atlas next to every PLY, including affine-to-sensor metadata
• render <stem>.render.png with the original NCore F-theta calibration, exact per-pixel world rays, exposure START/END poses, and rolling shutter
• add --merge --render-video to stream every original source exposure to an H.264 <stem>.render.mp4 at the source cadence without materializing the full ray sequence
• pin the public gsplat F-theta convergence fix and explicit 3DGUT world-ray path at a337c3dc
• fail explicitly for unsupported camera/distortion models, truncated --max-chunks, multi-sequence video input, missing libx264, or encoder failures

Standard 3DGS PLY has no environment-map or per-camera ISP fields, so existing PLY viewers continue to show the Gaussian foreground only. The sidecar preserves that non-PLY state for a sidecar-aware renderer, while the PNG/MP4 outputs provide immediately testable source-view renders.

Validation

• uv sync --frozen --extra render
• .venv/bin/pytest -q — 542 passed
• .venv/bin/ruff check --no-cache .
• uv lock --check
• git diff --check
• live CUDA exact-ray F-theta/3DGUT smoke with the pinned gsplat commit
• full public example clip, two chunks plus merge and all 599 source frames:

INSTANT_NUREC_RUN_ID=calibrated-sky-final .venv/bin/python run_inference.py \
  --model pa-front \
  --ncore-path demo_clip/clips/000da9de-0ee5-465a-9a2d-e7e91d3016bb/pai_000da9de-0ee5-465a-9a2d-e7e91d3016bb.json \
  --output-dir /tmp/instant-nurec-calibrated-final \
  --max-chunks 2 --merge --render-preview --render-video

The example completed successfully with 1,918,461 merged Gaussians. The MP4 is H.264 High/yuv420p, 784x448, exactly 599 frames, 29.9994 fps, and 19.967 s; a full decode found no corrupt, blank, or duplicate frames. Exact source/render comparisons at frames 0/299/598 measured static-feature displacement medians of 0.782 / 0.943 / 0.995 px (p90 1.630 / 2.043 / 2.487 px), with stable horizon and edge-to-edge geometry.

Renderer scope

The optional outputs are calibrated source-trajectory checks, not arbitrary novel-view renders. This public path currently supports NCore F-theta cameras without external windshield distortion and rejects other models rather than silently approximating them. Dynamic people/vehicles in the released public inference path remain subject to the model's current empty dynamic-layer output; that affects scene content, not camera calibration.