pybind/mgr/mds_autoscaler: deploy larger or smaller (RAM) MDS in response to MDS load
As a follow-up to #40929, it would be great to start deployment of MDS with smaller ones (like 4GB RAM). This would be helpful for multifs deployments with file systems scaling with different tenants.
The tricky part is growing the size of the MDS in response to load. That should be as simple as examining the running average of perf stats sent to the MGR. Look at the inodes loaded into cache and expired to see how much cache churn there is. Likewise, if the file system cools down, shrink the MDS down.
Actually growing/shrinking MDS size is done by adding a new MDS for that file system and removing the old one.