use array instead of valarray as underlying container for Tensor

Since we know all sizes at compile time we can eliminate all dynamic
allocations this way.