NaN boxing

[liquidev]

Features some optimizations that make values more compact.

A big bottleneck right now is the amount of destructors that kill branch prediction. I did some optimizations in places where the type of the value is known (namely, number operations) and std::mem::forget is called there to not run destructors when we know that the type of a value does not need any special work.

Here's a comparison between having a impl Drop for ValueImpl and not having it:

> hyperfine --warmup 5 'target/release/mica code/functions.mi' 'target/release/mica-nodestructors code/functions.mi'
Benchmark 1: target/release/mica code/functions.mi
  Time (mean ± σ):     344.4 ms ±   5.1 ms    [User: 342.1 ms, System: 1.9 ms]
  Range (min … max):   338.4 ms … 356.3 ms    10 runs
 
Benchmark 2: target/release/mica-nodestructors code/functions.mi
  Time (mean ± σ):     288.2 ms ±   1.5 ms    [User: 286.8 ms, System: 1.4 ms]
  Range (min … max):   286.6 ms … 291.2 ms    10 runs
 
Summary
  'target/release/mica-nodestructors code/functions.mi' ran
    1.20 ± 0.02 times faster than 'target/release/mica code/functions.mi'

Of course the faster version does not clean up any memory, which is bad.
With that I conclude that Mica needs garbage collection for improved performance.