Harden rDSN fault handling under fault injection

[linmajia]

Summary

Improve rDSN robustness under fault injection by converting several assertion/crash/no-progress paths into controlled failures, and by hardening simple_kv checkpoint persistence against partial writes.

Changes

• Improve network startup fault handling:

  • Propagate RPC provider startup failures instead of aborting through core assertions.
  • Handle socket/bind/listen/provider initialization failures more gracefully.

• Harden process startup under injected exceptions:

  • Catch startup-time exceptions in main().
  • Return controlled startup failure instead of crashing unexpectedly.

• Harden task-code and threadpool registration:

  • Prevent failed task-code registration from dereferencing invalid IDs.
  • Make registration paths tolerate allocation/container failures.
  • Avoid aborting during registration failure logging.

• Harden allocator and binary-reader fault paths:

  • Make callocator throw allocation failure instead of returning null to STL containers.
  • Handle corrupt replica app-info/binary-reader input without assertion crashes.

• Harden module loading:

  • Return controlled errors for module/symbol load failures instead of asserting.

• Harden simple_kv checkpoints:

  • Write checkpoints through a temporary file and publish via rename only after successful flush/close.
  • Validate checkpoint files during recovery.
  • Recover from the newest valid checkpoint instead of blindly trusting the newest file.
  • Return checkpoint corruption errors instead of asserting on malformed checkpoint data.
  • Validate learned checkpoint state before accessing state.files[0].

• Sync external submodules to merged master:

  • rDSN.dist.service: 4757232
  • rDSN.tools.hpc: 842ba7c

Validation

• Ran focused libfiu probes for the discovered failure classes.
• Confirmed the previously observed assertion/crash/no-progress cases were no longer reproduced in the focused probes.
• Full test suite ran successfully.