Fix High Memory Usage of RocksDB

[alexanderkiel]

We observe that the resident memory usage (RSS) of Blaze in certain cases is much higher as the sum of known constituents, the Java memory and the RocksDB block cache memory. Closer inspection revealed that the overhead depends on the size of the RocksDB block cache. The overhead is about 2 - 3 times the size of the block cache.

The root cause is that the default memory allocator of glibc shows high amounts of fragmentation of used in a multi core setting. Switching from glibc to jemalloc (#1014) solved the problem.