Decouple monarch-rdma from CUDA for hardware-neutral RDMA

[xiaopyyy]

Summary

monarch-rdma is now tightly coupled to CUDA/NVIDIA-specific components, which blocks Monarch from running on non-CUDA accelerators (e.g., Ascend NPUs). Decoupling RDMA from CUDA is required to achieve hardware-neutral accelerator support.

Key blocking dependencies

1. NVIDIA-specific infrastructure
   Reliance on GPUDirect RDMA, nvidia_peermem, and NVIDIA driver assumptions prevents RDMA use on non-NVIDIA systems.

2. CUDA-bound RDMA offloading
   rdmaxcel-sys implements critical RDMA operations (e.g., send_wqe, db_ring) as CUDA kernels, binding RDMA execution to the CUDA driver API.

3. CUDA-only PyTorch integration
   pytorch_segment_scanner depends on torch.cuda.memory._snapshot(), tying memory inspection and registration to the CUDA backend.

Goal / discussion

What would be the preferred approach to refactor monarch-rdma toward a hardware-neutral RDMA layer, for example, by introducing an accelerator-agnostic RDMA abstraction and isolating CUDA-specific optimizations behind optional backends?