feat: defaulting CPU offload + GPU OOM fix + viewer PLY export

[nvandamme]

Improvements for long-sequence inference stability and visualization flexibility: CPU offloading enabled by default, GPU OOM prevention via KV cache/memory management fixes, new PLY export in the point cloud viewer and some viewer UI fixes to avoid browser freeze when loading long-sequence inference results (+3000 frames). Additionally, pyproject.toml is now compatible with uv tooling as an alternative to conda.

---

demo.py — Default CPU Offload + CPU-Friendly Warmup

• Default --offload_to_cpu to True (was False) — per-frame predictions offload to CPU by default to cut GPU peak memory
• Add CUDA OOM warning when --no-offload_to_cpu is used with windowed mode or >512 frames
• CPU-friendly warmup: _warm_streaming() now moves only the warmup slices (not entire video) to device using non_blocking=True, preventing long videos from becoming persistent GPU residents before inference starts
• Images no longer moved to GPU before model pipeline starts — freed memory before inference
• Cross-platform path handling via pathlib.Path; deprecated Image.ROTATE_270 → Image.Transpose.ROTATE_270

lingbot_map/models/gct_stream_window.py — KV Cache + Memory Fixes

• Refactor aggregator/camera head access through local cast variables to ensure correct attribute resolution when the model is offloaded (fixes hasattr(self.aggregator, ...) failures on CPU-offloaded models)
• Delete unused tensor parameters (view_graphs, causal_graphs, ordered_video) from forward_aggregator immediately after receiving them to prevent GPU memory retention during streaming inference
• Fix _rollback_last_frame(): apply cast variables consistently across FlashInfer rollback, SDPA cache trimming, camera head rollback, and frame counter decrement — fixes corruption when rolling back on CPU-offloaded models
• Fix _execute_deferred_eviction(): same cast pattern for deferred eviction execution
• Add explicit memory deletion in windowed inference loop: del scale_images, del window_pred, del window_images after each window to prevent accumulation
• Replace old _align_and_stitch_windows() with incremental _merge_window_pred(): merges predictions on-the-fly instead of storing all windows then aligning, reducing peak memory from O(n×S) to O(window_size)
• Add alignment metadata (chunk_scales, chunk_transforms, alignment_mode="scaled") when merged window count > 1

lingbot_map/vis/point_cloud_viewer.py — Rendering Overhaul + PLY Export

• Replace deprecated matplotlib.cm (cm.get_cmap, matplotlib.cm.ScalarMappable) with modern from matplotlib import colormaps and colormaps.get_cmap()
• Input validation: raise ValueError when depth predictions are required but missing (use_point_map=False) and when confidence is needed for sky masking
• Replace per-frame point cloud handles (self.pc_handles, self.vis_pts_list) with single active handle (self.active_point_handle) + render caching — reduces browser scene graph size from O(n_frames) to O(history_window)
• New frame visibility system: "Current Frame Only" (history=0) / "Recent Frame History" (default=20 frames) instead of old 4D/3D toggle buttons
• Add gui_history_frames slider (0–200) and gui_max_render_points slider (50k–1M) for performance control on large sequences
• Start paused (Playing=False) to prevent immediate browser-side rendering of large point clouds
• PLY export: full .ply file export with normals, per-point confidence/frame-index/UV coordinates, camera metadata as extra elements, and scene alignment — triggered via new "Export PLY" button alongside existing GLB export under renamed "Export 3D" section

lingbot_map/vis/glb_export.py — Refactoring + PLY Export Functions

• Soft-import trimesh via importlib.util.find_spec() for cleaner dependency handling (still prints warning on missing)
• Extract _prepare_export_point_cloud(): shared point-cloud preparation logic used by both GLB and PLY exporters — handles filtering, sky masking, confidence thresholding, background masking, color coercion (_coerce_colors_to_uint8), and scene scale computation
• Add predictions_to_ply(): high-level export function mirroring predictions_to_glb() API with same filtering options (confidence threshold, frame filter, sky mask, background masking)
• Add save_point_cloud_to_ply(): writes binary little-endian PLY files supporting normals, per-point alpha, and arbitrary extra vertex properties/elements
• Add scene alignment helpers for standalone use: apply_scene_alignment_to_vertices(), apply_scene_alignment_to_directions(), get_scene_alignment_transform() — enables aligning vertices/normals outside of a trimesh Scene context (needed for PLY export)

Configuration Files

• .gitignore: Add uv.lock, *.ply, *.glb, *.7z, export_data.bin, lingbot-map-long.pt
• pyproject.toml: Python version pinned to ~=3.12.0; add torch>=2.12.1, torchvision>=0.27.1, flashinfer-python, pip>=26.1.2, flashinfer-cubin>=0.6.13; switch from onnxruntime to onnxruntime-gpu
• README.md: Added uv tooling alternative section mirroring conda step-by-step structure (uv venv --python 3.12, uv sync, optional --group vis / --group demo)