This is also a redesign of how all this code works, in particular
making `checkbounds(Bool, A, I...)` more first-class and splitting the
1d checking into `checkindex`.

This also fixes a serialize test that seems broken

I was trying to avoid adding another trait, but the reality is that this means less specialization for user-types (especially of "difficult" functions like index_shape), and may be the only way to write certain algorithms in a type-stable manner.

sub2ind is ambiguous when using unconventional indexing in 1-d, so handle all logic separately.
This fixes a problem with `show`ing OffsetVectors that cropped up after rebase.

Broadcast looks like it needs rewriting for jbfunctions. This starts on the process for these specific operations, but leaves the rest of the problem untouched.

One issue here is that I'm implicitly assuming that `parent(P)` returns the "inner" type requested by `allocate_for`. I haven't had to make that assumption yet.

…exing

Introduces promote_indices (is this a good name?) that picks one array to represent a list of arrays, as long as there is no inconsistency in array types.

Since the iterator is guaranteed to be inbounds, there is no reason to use a bounds-check. This leads to substantial performance improvements for certain array types.

Very worthwhile from a performance perspective, since bounds-checking is so performance-critical

I was hoping that a small amount of type instability would pay for itself later, but no such luck.

SubArrays have slow indices(S, d) but fast indices(S), so exploit this.