Short-circuiting expressions produce divergent types that propagate too far (#2700)

This is a fairly significant adjustment to how divergent types
propagate. Previously, every time a divergent type encountered another
type in propagation during inference, the divergence would cause type
erasure, losing all the information about the type calculated so far.
This means tuples, like `(<Int>, <Divergent>, <Double>)` would just
become `<Divergent>` which would allow it to unify with any type, even
tuples of a different arity. The same is true for arrays, where `[<Int>,
<Divergent>, <Double>]` would not produce an error even though the type
inference saw an Int and a Double that can't be in the same array. For
the fix, this continues to propagate the bool that indicates a type is
divergent without erasing the type. In specific cases, like the
boundaries of blocks in if-expressions or the block in a specialization,
the inference performs intentional type erasure similar to before. This
allows for short-circuit expressions to have the right effects while
still having better validation of types we can detect.

Fixes #2697

Author: swernli