Frontend 1/5 speedup: CPU reuse and upstream MegaLoc

[Zador-Pataki]

Changes

• Reuse prepared images, reconstruction data and dense match buffers.
• Speed up CPU selection and serialization.
• Use upstream MegaLoc attention, aggregation and faster model loading.

Impact

Shortened LaMAR HGE sequence: 589 images, 63 keyframes.

Frontend: 145.606 → 136.682 s (8.924 s saved, 6.13%).

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

Component measurement (61 keyframes)	Before (s)	After (s)	Saved (s)
Squared-distance calculations	1.719	0.321	1.398
Match selection	0.42699	0.01450	0.41249
Dense-buffer updates	0.41003	0.00147	0.40856
Metadata serialization	0.768	0.191	0.577
Adjacent-image loading	4.7519	2.9383	1.8136