v2026.9.30

Changelog
- General speed improvement in dataset creation
- Support splitting a large dataset into multiple parts, then train separately and merge
- Support natural language prompt for SAM 2.1 masking models
- Support [BiRefNet](https://github.com/zhengpeng7/birefnet) masking model
- Improved training ETA estimation, and added training VRAM forecasting and OOM warning
- Reduced chance of "VkResult -4" error in dataset creation and meshing
- Improved support for using telemetry information in dataset creation
- Support pen tool for masks
- Support using a separate fixed-area mask for each camera in dual-fisheye setups
- Fixed error loading PLY saved from a resumed training
- Fixed COLMAP dataset parsing error with relative paths
- Fixed video decoding issues with some drivers

Notes:
- This release carries difference in SfM intended to be improvement. Please submit an issue if you notice significant regression in accuracy or speed compared to previous versions.
- The default masking model has been changed from SAM 3 to SAM 2.1 Base+ with Grounding DINO, which provides 2-3x faster masking with slightly reduced accuracy. SAM 3 remains available as a non-default option.