Add GeoCalib-based focal estimation (per-image and per-folder) to sharp predict

[disk02]

Motivation

• Provide optional focal length estimation at inference time using GeoCalib to improve intrinsics when EXIF is missing or unreliable.
• Support both per-image calibration and an optional per-folder shared-intrinsics mode to amortize GeoCalib cost.
• Keep camera model fixed to pinhole and avoid writing calibration artifacts to disk.
• Fail gracefully to preserve existing behavior when GeoCalib is disabled or fails.

Description

• Added a GeoCalib adapter src/sharp/utils/geocalib.py that lazily imports geocalib, holds a single runner instance, and exposes calibrate_image and calibrate_folder returning a scalar f_px extracted robustly from varied GeoCalib return shapes.
• Extended src/sharp/cli/predict.py with --geocalib and --geocalib-per-folder flags, instantiated GeoCalibRunner on demand using CPU for mps devices, and integrated per-image or per-folder focal overrides while keeping original EXIF fallback.
• Added per-folder precomputation and in-memory cache (folder_fpx_cache) when --geocalib-per-folder is enabled, and logged concise info on computed folder focals and warnings when GeoCalib fails and EXIF fallback is used.
• Updated requirements.txt to add GeoCalib as a direct dependency via geocalib @ git+https://github.com/cvg/GeoCalib.git.

Testing

• No automated tests were executed as part of this change.

---

Codex Task