Changes:

- handle different CRS transparently

- don't build a spatial index on the selection layer. Instead
only use feature requests to fetch features which are within
the desired bounds, and rely on the presence of an appropriate
spatial index at the provider's backend. Otherwise, we force
every user of this algorithm to have a full iteration of the
source table, regardless of how large the table is. That means
that trying to select a set of addresses which fall within
a specific locality from a table which contains the addresses
for a whole state will FORCE every address in the state to
be initially read before any calculation begins. With this
change only those features within the bounding box of the
selected localities will ever be fetched from the provider,
resulting in huge speed improvements for the algorithm.

- use prepared geometries for the spatial relation tests.
This dramatically speeds up the algorithm in the case
where the intersection layer features cover multiple
features from the 'selection' layer.

- Add a 'select within current selection' mode

- Optimise feature requests for efficiency (especially
with respect to the 'disjoint' selection mode)