[+] Feat: Implement gdrcopy support for GPU maps

[Sy0307]

Implement #500.

• Enabled by: -DBPFTIME_ENABLE_GDRCOPY=ON
• Runtime controlled via: agent_config.enable_gpu_gdrcopy, with transparent fallback to cuMemcpyDtoH on failure.

Features

• Implement hybrid lookup policy: For BPF_MAP_TYPE_GPU_ARRAY_MAP and BPF_MAP_TYPE_PERGPUTD_ARRAY_MAP, elem_lookup tries copy_from_device_to_host_with_gdrcopy(...) first, otherwise falls back to cuMemcpyDtoH.
• Add per-key size threshold: agent_config.gpu_gdrcopy_max_per_key_bytes (default DEFAULT_GPU_GDRCOPY_MAX_PER_KEY_BYTES=4096); skip GDRCopy when per-key bytes exceed the threshold.
• Dynamic loading + fallback: Uses libgdrapi.so.2 + dlsym; missing libgdrapi.so or /dev/gdrdrv does not break functionality.

How to test

Build

cmake -S . -B build -G Ninja -DBPFTIME_ENABLE_CUDA_ATTACH=ON -DBPFTIME_CUDA_ROOT=/usr/local/cuda -DBPFTIME_ENABLE_GDRCOPY=ON
cmake --build build -j --target gpu_array_map_host_perf gpu_per_thread_array_map_host_perf

Run comparisons

Baseline:

./build/benchmark/gpu/host/gpu_array_map_host_perf --iters 50000 --max-entries 1024 --value-size 8 --gdrcopy 0

GDRCopy:

./build/benchmark/gpu/host/gpu_array_map_host_perf --iters 50000 --max-entries 1024 --value-size 8 --gdrcopy 1 --gdrcopy-max-per-key-bytes 4096

PERGPUTD baseline:

./build/benchmark/gpu/host/gpu_per_thread_array_map_host_perf --iters 50000 --max-entries 1024 --value-size 8 --thread-count 32 --gdrcopy 0

PERGPUTD GDRCopy:

./build/benchmark/gpu/host/gpu_per_thread_array_map_host_perf --iters 50000 --max-entries 1024 --value-size 8 --thread-count 32 --gdrcopy 1 --gdrcopy-max-per-key-bytes 4096

[Officeyutong]

Is the nullptr necessary? We may obtain the length of an array by std::size

[Officeyutong]

These declarations should be put in a seperate header file, I think