Fix member initialization order bug in Buffer constructor

`use_default_stream_as_comm_stream` was declared after `comm_stream` in the
class, but `comm_stream`'s initializer reads `use_default_stream_as_comm_stream`.
C++ initializes members in declaration order, so `comm_stream` was initialized
using an uninitialized bool — causing internode RDMA dispatch timeouts.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Implement returning num_recv_tokens_per_expert_as_cuda

Fix wheel relocatability: RPATH and nvshmem dependency

- setup.py: Use $ORIGIN-relative RPATH so deep_ep_cpp.so finds
  nvshmem from the pip-installed nvidia-nvshmem-cu12 package instead
  of hardcoding the build-time path.
- pyproject.toml: Add nvidia-nvshmem-cu12>=3.5.19 as a runtime
  dependency so pip pulls the correct nvshmem version.

Co-Authored-By: Claude Opus 4.6 <[email protected]>