Parallelised iradon

[jadball]

I was able to parallelise the iradon function.
This was achieved by distributing the interpolation calls across multiple processes.
I've also added a check to use the (supposedly faster) np.interp method instead of the scipy one if the desired interpolation is linear.
With 7 workers (seems to be the fastest for the shepp-logan phantom) and 3600 projections, this new parallelised version is around 6 times faster than before.

[jonwright]

Thanks! I will merge it like this. This code you have added is parallel over projections and it is doing a reduction into the final image (so a lot of summing on a single thread, or threaded by numpy?).
The original idea was to chop the output image into tiles (or rows) and then treat each one in a thread. So then there is no locking.
I think this can still improve this approach if each thread sums several projections and returns only one frame to be summed (e.g. do some of the reduction summation in parallel).