Improve VGC convergence

[zitongzhan]

Found likely bugs in InstantSfM’s view-graph calibration vs glomap/COLMAP:

• It optimized cameras that already had has_prior_focal_length=True; glomap keeps those fixed.
• It didn’t write calibrated state back like glomap:
  • cameras weren’t promoted to has_prior_focal_length=True
  • good pairs weren’t promoted to CALIBRATED
  • bad pairs weren’t marked DEGENERATE
  • E wasn’t recomputed from the refined calibration
• It filtered pair validity using unreverted degenerate focal solutions.
• It didn’t recompute fundamental from Essential for already-calibrated pairs before optimization.
• On ETH3D single-camera scenes, the same-camera optimization path converged worse than COLMAP; Use an exact bounded scalar solve against the same robust loss.

What was added:

• glomap-style calibration input preparation
• exact loss evaluation helpers
• prior-focal fixing behavior
• rejected-camera rollback before pair filtering
• calibrated/degenerate state updates
• ETH3D regression tests for:
  • no-intrinsics loss vs colmap view_graph_calibrator
  • gt-intrinsics loss vs colmap view_graph_calibrator
  • gt-intrinsics prior focal no-drift

Verification run:

• PYTHONPATH=/workspaces/InstantSfM INSTANTSFM_ETH3D_SCENES=botanical_garden,bridge,courtyard,facade,terrains pytest -q tests/test_view_graph_calibration.py -q