Frontend 4/5 speedup: DA3 inference acceleration

[Zador-Pataki]

Changes

• Speed up model construction and weight loading.
• Compile and save the backbones and depth heads.
• Batch four independent windows.
• Skip unused ray predictions while retaining depth, confidence and camera-derived intrinsics.
• Release finished weights without copying them to the CPU.

Impact

Shortened LaMAR HGE sequence: 589 images, 61 keyframes.

Frontend: 72.940 → 63.988 s (8.952 s saved, 12.27%).

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

Component measurement (61 keyframes)	Before (s)	After (s)	Saved (s)
Model construction and weight loading	≈9.60	≈4.41	≈5.19
Backbone compilation (estimated warm inference)	5.05	3.43	≈1.62
Skip unused ray predictions (estimated warm inference)	3.41	3.01	≈0.40
Depth-head compilation (estimated warm inference)	3.01	2.79	≈0.22
Four-window batching (estimated warm inference)	2.79	2.21	≈0.58
Release weights without CPU copy	1.974	0.168	1.806