Implement constant folding for casts to floating-point types.

[axch]

Caveat: Did not check that they agree bit-for-bit with the runtime behavior, but any discrepancies should be small.

Benefit:

• This ends up constant-folding the cast from the literal zero to the relevant floating-point type in the Add instance for floats.

• Therefore, that zero ends up being an Atom and not an expression when we build the BaseMonoid for adding floats, and so is inlined therein.

• Therefore, that zero is, in effect, rematerialized when doing AD.

• Therefore, an n by m by k tensor of those zeros is not stored by linearize of matmul, so the jvp-matmul benchmark runs some 30% faster. Of course, it is still storing two other tensors it shouldn't store, but that's for another time.

[axch]

Added some measure of consistency testing between run-time and compile-time casts. Unfortunately, it's hard to test many values, because compiling and running a million 1-instruction LLVM programs is much more expensive than executing a million LLVM instructions. That said, the tests did reveal two bugs (one already present and one introduced by me in the previous commit), which fixed.