* Fix float hash on 32 bit machines.

* Implement missing arithmetic operations.

* Implement missing comparisons.

* 0-d arrays should not be sequences.  Currently there are problems with using
  0-d arrays with PySequence_Fast.

* Consider reducing the amount of templated code by having more dtype-flexible
  code.

* Merge array_from_arraylike and matrix_from_arraylike

* Optimize readin_arraylike for elements that are tinyarrays.

* Implement concatenate.
