Intelligent broadcasting

Branching from #128, I'd be interested to see AxisArrays provide a method to automatically align AxisArrays for broadcasting in a manner parallel to what xarray does in Python.

For those of you who aren't familiar, the final line of this:

A = AxisArray(randn(10,15), :x, :t)
B = AxisArray(randn(10,20), :x, :y)
C = AxisArray(reshape(A, (10,15,1)) .* reshape(B, (10,1,20)), :x, :t, :y )

would become this:

C = A .* B

which is especially nice when one wants to abstract over the number shared & non-shared dimensions. Presently, the cleaner syntax here would fail due to the mismatch between 15 & 20. Additionally, even

C = A .* A

presently produces C as a plain array, not an AxisArray, which discards the axis information.

There are some gotchas: the ordering of the resulting dimensions could matter for performance reasons, and I don't think xarray has a way to specify how this is done. Additionally, if there are vectors associated to the axis names, then one has to decide if/how to test equality for same-name axes on different arrays, and what to do in the case of inequality. (xarray chose pointwise equality tests and silently drops any non-matching points along each axis. While I understand the motivation to imitate a database join, I've been bitten by floating-point comparisons in this context.)

Start with the broadcasting examples in this issue and the related discussion in #128, then inspect how AxisArrays currently handles broadcasting and preserves axis information. Define the dimension-ordering and same-name-axis equality rules before implementing automatic alignment; done means expressions such as A .* B align dimensions and return an AxisArray without discarding axis metadata.