Add world_up config flag so Y-up scenes render right-side up

[alexmgee]

Closes #4

What

Adds an opt-in world_up setting so splats stored with a Y-up world axis render upright instead of upside down. Default ("y-down", the standard 3DGS/COLMAP convention) leaves current behavior byte-for-byte unchanged.

• config.py - PipelineConfig.world_up: str = "y-down" plus a WORLD_UP_CHOICES constant (single source of truth for CLI + API validation, same as QUALITY_PRESETS)
• render_cameras.py - _build_poses() picks the camera up-vector from world_up ([0,1,0] for y-down, [0,-1,0] for y-up) and passes it to _lookat; the standalone CLI gains --world_up
• run_local.py / pipeline.py - --world_up {y-down,y-up} argparse flags, threaded through PipelineConfig.from_overrides exactly like score_threshold
• server.py - world_up form param on /api/v1/detect with a 400 on invalid values (mirrors the quality check), passed through the pipeline.py subprocess call, included in the job log lines
• README.md - parameter table row + the orientation warning now offers the flag as the remedy instead of re-exporting the file

Why

In the OpenCV-convention c2w built by _lookat, image-down for a level view equals the up argument, so it must point at the scene's floor. The hardcoded [0,1,0] is correct for Y-down scenes but renders every view upside down for Y-up scenes, and detections silently collapse to orientation-tolerant classes. On a real 5.08M-splat Y-up indoor scene (medium quality, 12-label prompt, identical gates), fixing orientation took raw OWLv2 detections from 663 to 1,255 and default-gate clusters from 3 (1 label) to 12 (6 labels) - full numbers and mechanism in #4.

Design notes

• Camera-side flip only. Re-orienting the splat data itself would require rotating quaternions and SH coefficients; flipping the up-vector is equivalent for rendering and risk-free.
• Elevation sweep mirrors (−55°…+40° becomes +55°…−40° relative to the scene) when the up-vector flips. Intended: coverage stays symmetric-ish either way for an opt-in flag.
• Camera placement is untouched - the density-aware sampler is orientation-agnostic.
• Web upload form deliberately not extended (API/CLI only), matching min_peak_score which is also API-only today. Happy to add a form control if you'd like it exposed there.
• Naming: went with world_up to describe the scene's up-axis convention; happy to rename (up_axis?) if you prefer.

Testing

• Unit: for level views the c2w image-down column is [0,1,0] under y-down and [0,-1,0] under y-up; default poses are identical to pre-change output; rotations stay orthonormal.
• Determinism: camera positions + all 90 medium-quality pose matrices recomputed on the 5.08M-splat test scene match the pre-refactor prototype run exactly (max abs diff 0.0).
• Server: FastAPI TestClient: invalid world_up → 400 naming the valid choices; y-up reaches the pipeline subprocess argv; omitted → default y-down.
• End-to-end: full medium runs on the real Y-up scene produced the before/after numbers above.