Star particles and gparts links over MPI

A big bundle of changes:

 - Introduce star particles. At the moment they are empty shells but they are ready to be used.
 - An `spart` is always linked to a `gpart` and will receive its gravitational forces from there.
 - A `gpart` now contains an enum containing information about whether it is linked to a `part`, to an `spart` or to nothing.
 - Added a more comprehensive test to check that the part<->gpart and spart<->gpart links are correct after every rebuild or MPI transaction.
 - `spart` get kicked and drifted and get their time-step computed. 
 - `spart` are read-in and written to snapshots (only if running with `-S` for now). 
 - Only the particle types that will be used are now read from ICs. Speeds-up the starting up process.
 - Corrects many incorrect communication of `gpart` over MPI. 

With this in we can run the EAGLE_25 over MPI with gpart and spart activated (e.g. `swift_mpi -s -g -S -t 16`). Note that this does not fix the remaining bug in #256 (Now fixed independently).

See merge request !310

When redistributing, the debugging checks need to verify that particles are sorted in the right node, not in the right cell.

Conflicts:
	src/Makefile.am
	src/gravity/Default/gravity_part.h

No sorted pair over mpi

Fix to #256. Do not use sorted pairs when the pair task is crossing over two MPI nodes. 



See merge request !312

Conflicts:
	src/runner_doiact.h

Matthieu Schaller authored Feb 02, 2017 and Peter W. Draper committed Feb 03, 2017

Added the missing SSE function in vector.h. Compiles but accuracy is not identical to AVX when running test27.