Improvements to maxParticleVelocity

[archermarx]

This PR improves the functionality of WarpXParticleContainer::maxParticleVelocity in two ways.

1. The function can now make use of the GPU for the reduction stage, just like meanParticleVelocity
2. The function now actually returns velocities, wheras before it returned momenta.

This PR also adds a version to MultiParticleContainer, so the maximum velocity among all particle species can be obtained.

As far as I can tell, this function is not used anywhere at the moment. I am however working on another PR which will make use of it, so I thought i would go ahead and send this off for feedback now.

[archermarx]

A question for those more experienced. Is the a reason we need to call OpenMP manually when we're not using GPU? My understanding is that AMReX should be able to use OpenMP just fine within the ParticleReduce function, but I'm basing a lot of this code on the getMeanParticleVelocity function above this, which adopts this strategy.

[archermarx]

Not sure why CUDA is failing to compile. Can compile with NVCC on my machine without errors.

[archermarx]

Thanks for the refactor. It seems all tests are green. Anything else needed before merge? @RemiLehe @ax3l

[RemiLehe]

Thanks for the PR @archermarx