ENH: Support multi-GPU computation

[timmens]

For SOTA models most PyLCM users solve and simulate the model on a server, which often has multiple GPUs available.

If multiple GPUs are available, PyLCM should be able to leverage this.

Where it could help

Multi-regime case:

In the multi-regime case, in a given period we compute the value function for each regime. These computations are independent and embarrassingly parallel and could be done on separate hardware.

Split-up state-space:

For very large state-spaces that cannot be reduced further using regimes, we could try to divide the state-space into as many pieces as there is hardware available.

[hmgaudecker]

Thanks -- we should leave this here, but I think the work should be better put into tranquilo... That is, it would ofc be nice to maximise the evaluation speed of one LCM run; but what matters in the end is how long it takes to get estimated parameters out.

[timmens]

That is a very good point, and I agree that this is not of priority right now.

However, I do think that there is value in being able to use optimizers that do not easily parallelize, and that implementing this feature should not be too complex.

This does not mean that I want to do it any time soon, but only that I believe it should be done at some point.

[hmgaudecker]

Linking #147 -- or am I missing something?