Add chunked evaluation mode #194
Description
Similar to ForwardDiff, we should offer a chunked evaluation mode. Instead of computing a single forward pass with the full index set, this would compute multiple forward passes with subsets of indices of fixed-size.
This would greatly reduce the required memory and could enable the use of e.g. custom stack-allocated 64-bit BitSets. Such statically sized patterns might also be the missing puzzle piece for GPU compatibility (see #67).