Instead of each float value being a heap-allocated Python object, use
unboxed C doubles to represent floats. This makes float operations
much faster, and this also significantly reduces memory use of floats
(when not stored in Python containers, which always use a boxed
representation).

Update IR to support float arithmetic and comparison ops, and float
literals.

Also add a few primitives corresponding to common math functions, such
as `math.sqrt`. These don't require any boxing or unboxing. (I will add
more of these in follow-up PRs.)

Use -113.0 as an overlapping error value for floats. This is similar
to native ints. Reuse much of the infrastructure we have to support
overlapping error values with native ints (e.g. various bitmaps).

Also improve support for negative float literals.

There are two backward compatibility breaks worth highlighting. First,
assigning an int value to a float variable is disallowed within mypyc,
since narrowing down to a different value representation is
inefficient and can lose precision. Second, information about float
subclasses is lost during unboxing.

This makes the bm_float benchmark about 5x faster and the
raytrace benchmark about 3x faster.

Closes mypyc/mypyc#966 (I'll create separate issues for remaining open
issues).