Convert remaining operators to use tensor pool

[robertknight]

#108 introduced a buffer pool to enable re-use of tensor buffers that are no longer needed, as outputs for future non-mutating operations. Only a subset of operators currently allocate from the pool. To get the full benefit the majority of ops need to be converted. Non-converted ops can be found by searching for operators which ignore their pool argument. This is indicated by having an underscore-prefixed _pool argument in their Operator::run implementation.

[robertknight]

This is now done except for a few operators where the output size is data-dependent (NonZero, Range) and the RNN operators. For the RNN ops, I need to complete #95 first to avoid conflicts.