Add f64 <-> i64/u64 conversions

[Shnatsel]

Part of #297

The combined diff is large but it's not as scary as it looks. This is what this PR does:

• Add 64-bit type conversions, with lowering to hardware instructions where available, and scalar fallback otherwise (482faea)
  • Faster AVX2 formulations are possible but that's for a follow-up PR to keep the size of this one manageable
• Optimize AVX-512 Float->Unsigned conversion for both 32 and 64 bits (39adc0e)
• The rest is generator refactoring for readability with no impact on the generated code, confined to its own separate commits

[danderson]

Reviewed commit by commit, which makes the refactors easy to reason about because they don't alter codegen output or tests.

[danderson]

Maybe Debug too? Only matters for println debugging though.

[Shnatsel]

It's an internal generator struct, we'll add it when it comes up

[danderson]

Does this literal survive compilation unchanged? In rust playground, loading it into an f64 rounds it up to 9223372036854776000.0

[Shnatsel]

That's a formatting artifact, the bitpattern is the same. Try this:

let x = 9223372036854775808.0_f64;

assert_eq!(x.to_bits(), 0x43e0_0000_0000_0000);
assert_eq!(x, 2.0_f64.powi(63));

I've added a comment on that, and then looked at it and added comments on the rest of the code as well because if you didn't write it it's probably completely inscrutable