Part of beam evaluation reduction: more clean-up and preparation

Changes:

• Generally avoiding some verbose constructs which could be done more concisely.
• Instead of passing around shared_ptr inside critical loops, references to patches are now stored as pointers. shared_ptr cause an extra synchronisation which I rather avoid in critical loops.
• This MR also avoids xtensor's += operator in a crucial place, as it turns out xtensor does this operation very slow. It is now integrated into the beam application step -- this not only avoids xtensor's +=, but also reduces the amount of times we need to go through the data.

I've heavily timed the changes and all changes had a positive effect on performance.