fix: prevent CPU OOM on large batches by reducing memory fragmentation

[pxl-pshr]

Summary

When processing large frame batches (1200+ frames at 1456×768), the node crashes with a CPU OOM error (DefaultCPUAllocator: not enough memory: you tried to allocate 16102195200 bytes) even on systems with plenty of RAM (tested with 191GB). The root cause is memory fragmentation from thousands of small tensor allocations accumulated over long-running inference loops (~59 minutes), not insufficient total memory.

Changes

• Pre-allocate contiguous output tensors (depth_out, conf_out, sky_out, ray_origin_out, ray_dir_out) instead of appending to Python lists — eliminates thousands of small allocations and the subsequent torch.cat() copies
• Free images_pt and normalized_images early to reduce peak memory (~32GB freed during inference)
• Lazily allocate ray tensors after first frame to handle variable model output resolutions
• Use in-place operations (clamp_(), zero_()) to avoid extra tensor copies
• Delete intermediate tensors (del depth_out, etc.) immediately after processing into final form
• Reconstruct RGB output from original images input instead of keeping images_pt alive for the entire loop
• Helper functions (process_tensor_to_image, process_tensor_to_mask) now accept both pre-allocated tensors and lists (backward compatible)

Test plan

• Tested with 50 frames — no regressions, outputs match expected formats
• Tested with 1200 frames at 1456×768 on 191GB RAM system — previously crashed, now completes successfully
• Verify all 11 outputs (depth, confidence, resized RGB, ray origin/direction, camera params, sky mask, gaussian PLY)
• Test with keep_model_size=True option
• Test with camera parameters input