datetime_tz isn't readily available, so it's a lot easier to just
package it within the nilmdb tree.

Adjusted setup.cfg so "python setup.py nosetests" now works correctly.
Also added a "test" alias so that "python setup.py test" works.

Now nilmdb.client, nilmdb.server, nilmdb.cmdline, and nilmdb.utils
are each their own modules, and there is a little bit more of a
logical separation between them.  Various changes scattered throughout
to fix naming (for example, nilmdb.nilmdb.NilmDBError is now
nilmdb.server.errors.NilmDBError).

Reduced usage of "from __future__ import absolute_import" as much
as possible.  It's still needed for the functions in the nilmdb/server
directory to be able to import the nilmdb module rather than the
nilmdb.py script.

This should hopefully ease future packaging a bit.

Inside the pycurl callback, we can't raise exceptions, because the
pycurl extension module will unconditionally print the exception
itself, and not pass it up to the caller.  Instead, we have the
callback return a value that tells curl to abort.  (-1 would be best,
in case we were given 0 bytes, but the extension doesn't support
that either).

This resolves the 'Exception("should die")' problem when interrupting
a streaming generator like stream_extract.

Relying on __del__ to clean up the thread isn't as reliable.

Curl will give an error if we call .setopt() while a .perform() is
in progress, for example if we try to do a stream_insert() while
in the middle of a stream_extract().  Move the setopt() to the
beginning of the get/put functions to ensure that we hit this
error before we mess with the URLs or anything else.

`nrows` may change if you restart the server; documented why this is
the case in the design.md file.  It's not a problem.

Passes tests, but doesn't really handle nrows (and removing partially
full files) correctly, when deleting near the end of the data.

Now supports "list /foo/bar" in addition to the older "list --path /foo/bar"