Offloading points and voxelmaps on GPU

[koide3]

This PR enables offloading point and voxelmap data of old submaps from GPU memory. It effectively reduces GPU memory consumption with a slight degradation of global optimization speed up to CPU-GPU data IO. A new parameter "gpu_memory_offload_mb" in "global_mapping_gpu.json" controls the threshold for GPU memory usage.

  // gpu_memory_offload_mb          : Threshold for GPU memory offloading in MB (0 = disable offloading).
  //                                : If GPU memory usage exceeds this value, old submap points and voxelmaps are offloaded to CPU memory.
  //                                : Note that the current implementation doesn't support offloading of factors and graphics data, and thus
  //                                : the actual GPU memory usage will be much larger than this value.
  //                                : (Setting this to a half or quarter of the total GPU memory would be a good choice).

TODO : Offloading in offline_viewer

[TokyoWarfare]

I've excelent news @koide3 ! The GPU offload works wonderfully!!

Previously, it would OOM at about a cuarter of the lenght of the road you see in the picture. As of now with a value of 4096 in the JSON it is workign wonders!!

this is all 14KM mapped at max resolution

Dataset:
#271

Dense config with GPU offload set at 4096 (GPU--->3090Ti)
02_Dense_gpu_offload.zip

[TokyoWarfare]

sadly, as spected, OOM occurs in the offline viewer:

[2025-12-26 20:09:15.382] [viewer] [info] Use config from /data/MyHorizonBags/ROS2/00_Graphs/BAG_04_B112_HD/dump/config [2025-12-26 20:09:15.382] [viewer] [info] Export classes from libimu_validator.so [2025-12-26 20:09:18.135] [viewer] [info] enable_optimization=true [2025-12-26 20:09:18.256] [global] [info] Load submaps (session_id=0) warning: cudaErrorMemoryAllocation : out of memory warning: cudaErrorMemoryAllocation : out of memory warning: cudaErrorInvalidValue : invalid argument warning: cudaErrorIllegalAddress : an illegal memory access was encountered warning: cudaErrorIllegalAddress

I hope this GPU offload feature can be finished as would be a huge step for mapping.