Release 0.6.0

Features:

• Introducing a brand-new architecture design based on Ray, which can be used to solve deep learning problems in all scenarios.
• Introduce new concept of Node Group and its corresponding implementation to support existing capabilities such as fault tolerance on super node. Note: requires scheduling support from the Kubernetes side).
• New pre-check on device-memory using.
• Support Moore-Threads-GPU device. @zhwentao
• Support controller/master image with arm64 architecture.
• Support NUMA affinity using.
• Support specified node exit reason expansion which can be resolved from pod info by api-server.

BugFix and Enhancement:

• Refactor dependency requirements.
• Update pre-commit tools for development.
• Enhance node object generation when node relaunching.
• Enhancement on dealing with orphan process.
• Add validation for server-client using UID variable to avoid possible communication interaction contamination in large cluster.
• Remove the mandatory dependency on pynvml from the code. Thx advice from @yifeng-x #1651 #1647
• Add a more robust reclamation mechanism for process leaks in extreme scenarios.
• Fixed node failure reason reporting.
• Enhancement on rendezvous event.
• Enhancement for Ascend device.
• Other fixes and enhancements.