Implemented sampler on Arbitrary grid and fixes for cartesian one

[shkodm]

On Arbitrary grid:

• Moved sample to the lattice base
• Implemented executor to map cartesian coordinate into local id by an array scan with a tolerance. This is the only solution I think that avoid additional array allocations
• Otherwise the functions are similar to the cartesian ones
• To not search for index on every iteration, cached the point index inside the sample points attribute (not that much extra memory, should be ok)

On Cartesian grid

• Fixed a bug that on multiple GPUs the sample point location was never converted to the local coordinate frame, so was returning wrong results (because intersect does the intersection of bounding boxes, not the conversion into coordinate frames). Fix by offsetting the point into the local coordinate system

Tested by:

• Print local ids of different sampling points across 1 and 2 gpus
• Running sampler with several points on 1 GPU, and 2 GPUs and comparing results for both cartesian and arb lattice (to make sure the distribution of points works) - this is how I found the problem with the cartesian lattice, I think Lisa saw something like that previously
• Comparing sampler results between cartesian and arbitrary lattice on 1 and 2 gpus - no difference

Bounding Box changes are the same as in #540

[llaniewski]

No node type is used here.

[shkodm]

Had the same issue with override virtual functions as in #540 (comment). Should be fixed now