enhance memory management `csrc/filtering/permutohedral`?

[wyli]

it would be nicer to replace the c arrays by c++ objects (std::vector or std::array when possible). The current memory management seems brittle.

Originally posted by @tvercaut in #3698 (comment)

[wyli]

looks like this will copy data twice in some cases,

MONAI/monai/csrc/filtering/permutohedral/permutohedral.cpp

Line 33 in 927483b

torch::Tensor data = input.clone().movedim(1, -1).contiguous();

might be slightly improved by something like input.movedim(1, -1).clone(memory_format=torch.contiguous_format);? cc @charliebudd