This would be extremely useful! Please reenable the sampling profiler on one (or subset) of your pods but use the agent snapshot provided above. Set the JVM args to provide the most detailed heap dump if and when it crashes.
If this agent snapshot fixes the issue - great. If not, please try again to analyse your heap dump as you did, to figure out exactly which field/collection of the SamplingProfiler instance is holding these huge CallTree$Root objects (assuming you cannot provide the heap dump itself).
Thanks!