I believe there is a small race in the `getMDBEnv`: if the database
file does not exist when we first try to get the file metadata, we
acquire the lock then create a new `MDBEnv` and store it in the map.
But what happens if a different thread created the database between
our first check and the call to `MDBEnv`? I believe we would create
a new environment and override the existing entry in the map, bypassing
the check.
This commit introduces a second check right after acquiring the lock
to prevent that.