Frontend 2/5 speedup: Persistent RoMa compilation

[Zador-Pataki]

Changes

• Save RoMa model state and compiled programs to skip construction and graph capture in new processes.
• Make RoMa wrappers serializable with PyTorch while retaining descriptor normalization.
• Validate cached programs with identities and checksums; lock and publish entries atomically.
• Isolate internal cache lookup during compilation to prevent PyTorch's seeded cache-key collision.

Older PyTorch versions without program persistence use ordinary compilation.

Impact

Shortened LaMAR HGE sequence: 589 images, 63 keyframes.

Frontend: 136.682 → 103.127 s (33.554 s saved, 24.55%).

Median of three fresh local RTX 4090 runs with warm caches.