Accelerate PF and FGO GPU paths

[rsasaki0109]

Summary

• finish PF reductions for ESS, spread, estimate, and systematic-resampling normalization on device
• add an opt-in/auto cuSOLVER dense Cholesky path for basic and velocity/IMU FGO with guarded CPU fallback
• wire scoped PF-FGO experiment options, add a reproducible GPU benchmark, and document real-data/fault validation

Performance

• PF at 1,000,000 particles: 1.07x–1.29x faster for the optimized reductions/resampling
• FGO state size 800: 8.63x faster; state size 2000: 37.73x faster
• CPU/GPU maximum state difference: 4.55e-13

Validation

• CUDA Release builds: _gnss_gpu, _gnss_gpu_pf_device
• GPU-forced FGO: 73 passed
• selected PF/FGO, realtime, NLOS, cycle-slip, outage, and fault suites: 131 passed, 1 skipped
• CI-equivalent Python smoke: 329 passed, 1 skipped
• CI-equivalent wrapper smoke: 77 passed
• Tokyo/Nagoya existing-data replay: semantic parity; maximum ESS-ratio delta 9.3e-13
• git diff --check and commit-message lint pass

Default compatibility is preserved: the submission route remains CPU/off unless explicitly enabled, while evaluation may select auto.