feat: Low-VRAM mode for GPUs with <=12GB

[jashshah999]

Summary

Adds --low_vram flag that auto-configures VGGT-SLAM for GPUs with limited VRAM (8-16GB). Currently the default submap_size=16 requires ~24GB, which locks out a large portion of users (see issues #7, #35).

Changes:

• New --low_vram flag: auto-detects GPU VRAM and sets appropriate submap_size (4 for 8GB, 6 for 12GB, 10 for 16GB)
• New --checkpoint_inference flag: enables gradient checkpointing during eval (recomputes activations instead of storing them, ~40% VRAM savings)
• New --sequential_heads flag: runs camera_head and depth_head one at a time instead of holding both sets of intermediates
• Added torch.cuda.empty_cache() between submap inferences to reclaim fragmented memory

Usage:

# Auto-configure everything based on your GPU
python main.py --image_folder path/to/images --low_vram

# Or manually tune:
python main.py --image_folder path/to/images --submap_size 6 --checkpoint_inference --sequential_heads

Note: The --checkpoint_inference and --sequential_heads flags require companion changes in VGGT_SPARK (the aggregator and model forward pass). I'll open a companion PR there. Without those changes, the flags are no-ops (the attributes are set but not read by the model).

Estimated VRAM usage

GPU	submap_size	Checkpointing	Estimated peak VRAM
RTX 4090 (24GB)	16 (default)	Off	~20GB
RTX 3090 (24GB)	16 (default)	Off	~20GB
RTX 4070 (12GB)	6	On	~10GB
RTX 3060 (12GB)	6	On	~10GB
RTX 4060 (8GB)	4	On	~7GB

Test plan

• Run on RTX 3060 12GB with --low_vram (submap_size=6)
• Verify poses match baseline within tolerance on TUM office dataset
• Check that checkpointing produces identical outputs to non-checkpointed version
• Measure actual peak VRAM with torch.cuda.max_memory_allocated()

[jashshah999]

Benchmark Results (NVIDIA L4, 24GB VRAM)

Tested on the office_loop dataset (473 images, 208 keyframes selected):

VRAM usage at full resolution (518x518) by submap_size:

submap_size	Peak VRAM (inference only)	Total with model
4	3.78 GB	~6.1 GB
8	5.07 GB	~7.4 GB
12	5.59 GB	~7.9 GB
16	6.12 GB	~8.5 GB

This means:

• 8GB GPU: submap_size=4 works (6.1GB peak) ✓
• 12GB GPU: submap_size=8 works (7.4GB peak) ✓
• 16GB GPU: submap_size=12 works (7.9GB peak) ✓

Performance (submap_size=8, no loop closure):

• 208 frames processed in 55.7s
• 3.7 FPS average
• VGGT inference: ~1.06s per 9-frame submap

Note: The images in office_loop are 294x518. Full 518x518 images use slightly more memory as shown above.