fix(utils): make find_node_ip return a usable address on loopback-only hosts

[li-lizhe]

Summary

find_node_ip() returns main_ip or virtual_ip, which is None on a host where no interface other than the loopback one exposes an IPv4 address — for example a process started with docker run --network none.

return main_ip or virtual_ip

RayHelper assigns that value to MASTER_ADDR:

env_vars['MASTER_ADDR'] = ip

None is not a valid value there, so the failure surfaces inside Ray's env_vars validation with an error that does not point back to the address lookup.

This PR makes the final fallback 127.0.0.1, which matches the existing intent of falling back to a virtual interface, and leaves the preference order (non-virtual interface → virtual interface) unchanged.

Verification

Pure-Python behaviour, reproduced before/after by loading src/twinkle/utils/network.py from a file path (with a fake psutil) and calling the real find_node_ip():

########## BEFORE (upstream main) ##########
  only loopback (docker --network none)      -> find_node_ip()=None  MASTER_ADDR=None  -> Ray env_vars TypeError (not a str)
  loopback + docker0                         -> find_node_ip()='172.17.0.1'  MASTER_ADDR='172.17.0.1'
  normal host                                -> find_node_ip()='10.0.0.5'  MASTER_ADDR='10.0.0.5'
########## AFTER (patched) ##########
  only loopback (docker --network none)      -> find_node_ip()='127.0.0.1'  MASTER_ADDR='127.0.0.1'
  loopback + docker0                         -> find_node_ip()='172.17.0.1'  MASTER_ADDR='172.17.0.1'
  normal host                                -> find_node_ip()='10.0.0.5'  MASTER_ADDR='10.0.0.5'

The first two cases are unchanged; only the loopback-only case changes. A regression test covering all four cases is added in tests/utils/test_network.py.

Not tested

• No Ray cluster was started, so the exact downstream exception inside Ray was not reproduced; only the None value handed to MASTER_ADDR is demonstrated.

Fixes #302