We were still calculating the maximum number of rows correctly,
so the extra data was really extra and would get re-written to the
beginning of the subsequent file.

The only case in which this would lead to database issues is if the
very last file was lengthened incorrectly, and the "nrows" calculation
would therefore be wrong when the database was reopened.  Still, even
in that case, it should just leave a small gap in the data, not cause
any errors.

ENOENT might show up if we're actively deleting files in the nilmdb
thread while trying to read available space from e.g. the server
thread.

Instead of returning a single number at the end of N nilmdb calls, we
now use a generator that returns one line of text every time there's a
new count of rows removed.  This ensures that the connection will stay
alive for very long removals.

This lets you do something like "nilmtool extract | head" without
triggering backtraces.

This includes both client.stream_insert_numpy and
client.stream_insert_numpy_context().  The test code is based on
similar test code for client.stream_insert_context, so it should be
fairly complete.

Still allow installation without it, though.

This is a subclass of nilmdb.client.client.Client that adds numpy
specific routines, which should be a lot faster.