feat(gpu-remoting): add server-side CUDA callback handlers, storage data

[madd2014]

pipeline, and validation test samples
Add three core modules for the GPU virtualization remoting service:

• server: Implement RPC-based server endpoint with ~150 CUDA/cuBLAS/cuDNN/NCCL API callback handlers, virtual-to-real address mapping, checkpoint backup/recovery, and multi-client scheduling support

  - storage: Build the remote data feeding pipeline with
  

  shared-memory ring queue
  for zero-copy batch transfer to GPU, PyTorch
  DataLoader-compatible transparent API
  (FlexGVirtDataLoader), and 3D medical imaging preprocessing
  utilities

    - sample: Add comprehensive CUDA validation test suite
      covering memory management,
        kernel launch, stream/event/device APIs, CUDA Graph,
    multi-threading, cuBLAS, cuBLASLt, cuDNN, NCCL multi-node communication, and performance benchmarks