Issue using memory resource in `available_node_types`

First of all, thanks for a great project :+1:

I’m having some issues with Memory Aware Scheduling on a cluster with multiple node types.

I’m not familiar with the ray code, but the issue seems to be that the memory part of the resources attribute in available_node_types is being interpreted with different units in different parts of the ray code.

This is my understanding of how it works now:
python/ray/autoscaler/_private/ uses the 50 MB unit, while python/ray/ which gets it through RAY_OVERRIDE_RESOURCES interprets it as in bytes.
The RAY_OVERRIDE_RESOURCES is set in python/ray/autoscaler/_private/ from the node_resources attribute of NodeUpdater.

Thanks in advance for any help.