Allow a different eltype only for the coordinates

[efaulhaber]

This PR adds a keyword coordinates_eltype to the RectangularTank, RectangularShape and SphereShape, which defaults to Float64.
This will make the coordinates Float64, while everything else is still Float32.
For large numbers of particles, the particle spacing becomes very small relative to the domain size, which create larger errors in the distances between neighboring particles (relative to the particle spacing). With Float32 coordinates, these errors can easily grow to a point where they cause artifacts in the simulation.

This can easily be demonstrated by running the 2D dam break and passing min_coordinates = (1000.0, 1000.0) to the RectangularTank. With Float32, this yields:

With this PR and Float64 coordinates:

Here is a benchmark of a full dam break simulation and of just the fluid-fluid interaction kernel with 320k particles on an Nvidia RTX A4500.

Full simulation	Time	Relative
Float64	213 s	13.6×
Float32	15.7 s	1×
Float32 + Float64 coordinates	21.1 s	1.34×

fluid-fluid interact	Time	Relative
All Float64	16.823 ms	20×
All Float32	839.356 μs	1×
Float32 + Float64 coordinates	1.278 ms	1.5×

[svchb]

/run-gpu-tests

[svchb]

/run-gpu-tests

[svchb]

/run-gpu-tests

[efaulhaber]

/run-gpu-tests

[svchb]

/run-gpu-tests

[efaulhaber]

/run-gpu-tests

[svchb]

/run-gpu-tests

[LasNikas]

I'm, wondering if we want the ELTYPE and coordinates eltype info in the show box for Semidiscretization, too?

[svchb]

/run-gpu-tests