Frontend 3/5 speedup: RoMa inference reuse and fused correlation

[Zador-Pataki]

Changes

• Reuse RoMa features and projections across keyframe, tracking and loop-closure pairs; projection reuse reduces retained feature memory by ≈35%.
• Transfer images to the GPU only when their features need extracting, saving ≈2.17 GB of keyframe transfers.
• Pad the final batch to avoid compiling another input shape.
• Fuse local correlation.

Impact

Shortened LaMAR HGE sequence: 589 images; 63 → 61 selected keyframes.

Frontend: 103.127 → 72.940 s (30.187 s saved, 29.27%).

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

Component measurement (61 keyframes)	Before (s)	After (s)	Saved (s)
Keyframe feature reuse (replay)	≈18.43	≈17.42	≈1.01
Tracking feature reuse (replay)	≈29.18	≈24.12	≈5.06
Loop-closure feature reuse (replay)	≈7.45	≈6.43	≈1.02
Final-batch padding, when the tail triggers compilation	14.181	0.260	13.921
Fused correlation (estimated matcher workload)	41.58	23.88	≈17.70
Projection reuse (estimate, excluding projection of new views)	23.78	20.82	≈2.96