Optimize Model decoding

[dilation]

#87 improved the performance of Model decoding (i.e., the ModelRepr.toModel method). However it still seems slow.

Rudimentary profiling points to:

• the Eval interpreter
• lazy val initializers (some of them in Eval.Later)

We should identify whether the cause is synchronization of lazy vals, the Eval interpreter itself, or both (or something else).

Optimization ideas for the two cases:

• Avoid Eval/State
  • would a simpler State help?
• Thread-unsafe laziness, followed by safe publication (after decoding)
  • See some discussion about safe publication: https://github.com/scala/collection-strawman/issues/50
  • Some more reading: https://shipilev.net/blog/2014/safe-public-construction/