Add CUDA bundle adjustment support

[Kusum-mrv20]

Summary

• Add CUDA Sparse Levenberg-Marquardt bundle adjustment support through GTSAM PR #2761.
• Expose PCG/CUDSS settings, fallback diagnostics, and a runnable CUDA config.
• Add focused wrapper tests and documentation.
• Add a slow Gerrard Hall CPU vs CUDA BA timing test for review.

Validation

• uv run pytest tests/bundle/test_bundle_adjustment.py -q: 15 passed
• Real GPU BA on bundled Dubrovnik examples: CUDA device backend, converged in 4 iterations.
• dubrovnik-3-7-pre: error 691.0550 -> 0.3675
• dubrovnik-3-7-18-pre: error 688.2935 -> 0.0092

Gerrard Hall timing (for review)

CPU vs CUDA Sparse LM on the Gerrard Hall COLMAP sparse model (~100 cameras, ~43k tracks), measured with:

bash .github/scripts/download_single_benchmark.sh gerrard-hall-100 wget
uv run pytest tests/bundle/test_cuda_ba_timing.py -m "slow and cuda" -s --no-cov

Representative local result on an RTX 3080 (PCG backend, reprojection-only graph, max_iterations=50):

Backend	Optimize time	Final error	Notes
CPU LM	~99 s	21179.56	gtsam.LevenbergMarquardtOptimizer
CUDA Sparse LM	~13 s	21153.36	Device backend, no CPU fallback
Speedup	~7.7×	within ~0.2%	timed via _last_optimization_duration_sec

[Kusum-mrv20]

@kathirgounder can you please provide some initial reviews on these changes, have done sanity tests on small datasets, but can do further tests after first reviews

[Kusum-mrv20]

Hi Prof. @dellaert
Got initial review from Kathir, requesting your review on this PR