Fix numpy and XGMI 1-hop detection

[mawong-amd]

Temporarily fix numpy < 2.0.0 as Numpy 2.0 breaks ROCm PyTorch. Counterpart to vllm-project#5582

Fix XGMI 1-hop detection: previous version has the following problems

1. Ignores the device_ids passed in.
2. Each device only checks to see if it's 1-hop XGMI connected to all other devices. Instead, each device should check that all devices are 1-hop XGMI connected to all other devices. This prevents the odd case where some devices are 1-hop XGMI connected to all other devices, but others are not, which would result in not every device enabling custom_all_reduce and hence deadlock.