Mitigate risks of OOM with SoftReference cache

[devinrsmith]

A deep dive into #5516 exposed a situation where the JVM was getting OOM errors while there was plenty of reclaimable heap (99.9% of heap was taken up by reclaimable SoftReference'd objects). That fix itself should greatly reduce the risk of this error in this exact situation, but fundamentally other code paths can still hit JNI critical sections and cause "Retried waiting for GCLocker too often allocating" and OOMs. See https://github.com/devinrsmith/GCLockerTooOftenAllocating for a more detailed reproduction.

In the meantime, we can mitigate the risks by increasing -XX:GCLockerRetryAllocationCount above the default of 2.