[P2P] Multi-backend Support and Auto-detection

[derekwin]

• Bundle the native shared libraries into the UCCL Python package, eliminating the need for users to manually set library paths (e.g., when using Nix or similar environments).

• Enable automatic backend selection based on hardware capabilities:

  • Use the RDMA engine on machines with RDMA support,
  • Fall back to EFA backend on AWS instances with EFA NICs,
  • Use NCCL-TCPX on systems without specialized networking hardware.

• Add pybind11 support for all communication backends (not just the RDMA engine), so they can be used via the uccl.p2p Python API.

[praveingk]

This is good 👍🏽