Multi threaded L0 -> L1 compaction

Currently L0 -> L1 compactions are almost always single threaded because every L0 sstable will overlap with all L1 sstables. To improve this, we should range-split the input sstables in a configurable amount of parts and then use multiple threads to write out the results. This is similar to the max_subcompactions option in RocksDB: https://github.com/facebook/rocksdb/wiki/Leveled-Compaction