Adaptive chunking (#58)

* try holding a write lock for less time while still keeping it safe

* maybe this'll work

* a bit more logging

* attempt to handle multiple changes at once

* try to fix spammy rtt warning + try to dedup changes

* bomb when errors are encountered inserting multiple changes

* seqs start at 0

* disabling some optimizations

* this is weird, but testing if it's the issue

* why this would make things unreliable is beyond me

* one transaction = one db version, forgot about that...

* attempt to get rid of unique constraint errors

* better query for counting buffered changes, if count drops to zero then we should update the gauge

* more logs

* attempt to clean up mess

* continue loop afterwards

* clear buffered meta when applying complete change

* shorter sync backoff