A more reasonable way to obtain RDMA devices

[specture724]

resolve: #35

_parse_NCCL_IB_HCA method added, supporting exact match( "=" ), exclude( "^" ) prefix. Port specifications( ":" ) is not supported because mooncake transfer engine doesn't support it yet.

https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html#id8

[Copilot]

Pull Request Overview

This PR introduces a new NCCL_IB_HCA parser to support more sophisticated RDMA device selection, including exact match (=), exclude (^) prefix, and port specifications (:) syntax as documented in NVIDIA's NCCL user guide.

Key changes:

• Replaced simple substring matching with comprehensive parsing logic in _parse_NCCL_IB_HCA()
• Added helper function _resolve_device_specs() to handle device name resolution and port specifications
• Implemented comprehensive test coverage for various parsing scenarios

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
checkpoint_engine/ps.py	Refactored RDMA device selection logic by replacing inline filtering with dedicated parsing functions that support NCCL_IB_HCA syntax
tests/test_rdma_parser.py	Added comprehensive test suite covering basic parsing, pattern matching, error cases, and device allocation scenarios

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[weixiao-huang]

cc @whybeyoung plz have a look