Possibility to set custom precision in simulation

[federico-carrara]

Description

Related to #94, when simulating batches of samples at the same time, memory consumption quickly becomes a bottleneck (e.g., consider large 3D images with multiple channels). In this case, the user may want to trade off precision in the computation with the possibility of simulating larger batches at once.

For this reason, one cool feature would be to let the user select the desired precision (mainly single or double, I doubt that all the operations are supported for float16).

Potential problems

I foresee that guaranteeing compatibility with all the currently supported backends can be troublesome and require some thinking/work...

[federico-carrara]

Update: I honestly missed the float_dtype parameter in Simulation.settings! However, in my use case, it seems that the simulation doesn't stick to the chosen type, e.g., even if I start with float32 dtype, at some point the intermediate arrays are cast to float64. I will check which of the operations cause casting to double precision !

[tlambert03]

yeah, thanks. sounds like a bug. that is the intended place for that setting