skully: fix various bugs

[zerbina]

Summary

Fix various bugs with skully. For the most part, they were about
incorrect code being emitted, resulting in misbehaving programs.

Details

• fix size/alignment of embedded records (has no observable effect on
  compilation, at the moment)
• fix nil not being handled in constants, causing a crash
• fix incorrect code being emitted for closure equality tests
• fix pointer-location default initialization using an integer zero
• fix zero being used instead of Nil in various places
• fix wrong conversion operators being used for integer conversion
• fix integer extension conversion operators not being emitted
• fix ord not being translated to a conversion
• fix code emitted for mDestroy accessing the wrong field
• fix wrong code being emitted for cast where at least one operand is
  a float or imported type

[zerbina]

Most of these problems were found as part of #146, thanks to the stronger type system (when compared to the VM) of LLVM.

Most of the issues would have also been caught by a type checking pass that skully runs on its output, at least when used in conjunction with building phy with skully. As evidenced by the amount of issues, just running a skully-compiled phy does not provide enough coverage.

Testing skully via unit-like tests is hard, due the MIR + magics having a very large surface area -- there are also no NimSkull-provided facilities for constructing standalone MIR modules (which would be a requirement for testing). Running NimSkull's test suite with skully is also not possible at the moment, due to skully not using the nim CLI.

Ultimately, skully shouldn't have do the translation work it has do now; most of it should be done by the NimSkull compiler. I'm still looking into ways to upstream parts of skully.