Optimize Display.imageData()

[fengb]

The pixel translation is actually a pretty strong bottleneck — 33% of execution time.

Thoughts:

1. Mutate canvas directly — is this possible?
2. WebWorker
3. Separate wasm for translation — might not be efficient to copy the data around

[fengb]

This should be deferred until after GBC support since that will have a radically different layout for pixel values.

Edit: this change has been applied so we can optimize now!

[fengb]

LLVM gives us this output: https://godbolt.org/z/YXLf4D. Translated to Javascript:

function foo(r0) {
  var r1 = 255;
  var r2 = r0 & 992;
  r1 = r1 & (r0 << 3);
  r0 = r0 & 31744;
  r1 = r1 | (r2 << 6);
  r0 = r1 | (r0 << 9);
  r0 = r0 | -16777216;
  return r0;
}

function simplified(raw) {
  var r0 = 31744 & raw;
  var r1 = 255 & (raw << 3);
  var r2 = 992 & raw;

  return -16777216 | (r0 << 9) | r1 | (r2 << 6);
}

[fengb]

Applying the LLVM voodoo led to 10-20% performance boost: f29bca4#diff-c2b923b8185dd06e822dec6c545dae85R100