Fix RDMA chunking for oversized multi-SGE WRs

[maning00]

Summary

• Fix RDMA chunk expansion so oversized multi-SGE WRs are split before ibv_post_send.
• Carry the local ibverbs max_msg_sz capability on RdmaEndpoint without serializing it to peers.
• Add coverage for multi-SGE chunk geometry, max message limits, SGE boundaries, and disabled-chunking failure behavior.

Background

A large batch can be merged into a multi-SGE WR whose total message length exceeds the local provider max_msg_sz. The old chunker only split oversized single-SGE WRs, so these multi-SGE WRs could bypass chunking and be rejected by the verbs provider. This keeps the merge fast path for small WRs while enforcing the local message-size limit when transfer chunking is enabled.

Validation

• BUILD_TESTS=1 pip install .
• ./build/tests/cpp/test_engine