compact: consider value separation during heuristic for sstable splitting

Currently, during compaction, we split sstables based on `targetFileSize`. When value separation is enabled, the size of an sstable's blob references is included in this calculation:

https://github.com/annrpom/pebble/blob/edebc31de6b06c32fa37279927d6c7f2c476cfed/internal/compact/run.go#L283-L287

In a one-off test with an LSM containing ~4 TiB of separated values, the average sstable size seemed unexpectedly small (~60 KiB):

We should update our sstable-splitting logic to better account for aggressive value separation (at the very least) to see if we can bias less towards smaller sstables

Jira issue: PEBBLE-1142