(Fairly) fast iteration for ReshapedArrays

[timholy]

For some situations, this appears to solve the majority of performance problems with ReshapedArrays, even for complicated iteration patterns like iterating over the columns of a ReshapedArray. For the latter, the main remaining limitation is probably the fact that we can't construct allocation-free containers until they can be stored on the stack. So for ReshapedArrays with small first dimensions, it will not be so good.

Given those caveats, performance is already pretty good. Using the test at the end of test/dense.jl,

julia> @time sum_cols_slow!(S1, R);   # this uses traditional CartesianIndexing
  0.180041 seconds (4 allocations: 160 bytes)

julia> @time sum_cols_fast!(S1, R);    # this uses fancy indexing over each column
  0.074804 seconds (34 allocations: 960 bytes)

That's within a factor of 2 of generic LinearSlow iteration; would be nice to close that, too, but this is at least progress.