Optimize the ABI boundary of wasm-compiled proc macros

[alexcrichton]

This is a culmination of the following set of ideas:

• https://github.com/dtolnay/watt/issues/2#issuecomment-547643870 - investigation of current performance
• https://github.com/dtolnay/watt/issues/2#issuecomment-547662596 - idea of how to improve performance (this PR)
• https://github.com/dtolnay/watt/issues/2#issuecomment-548001322 - idea of how to further improve performance over previous idea (also this PR)
• https://github.com/dtolnay/watt/issues/2#issuecomment-548035170 - initial measurement of this PR

I figured these are all large enough changes that this'll want to eventually be bumped to 0.2, so I've started pruning APIs exposed by watt in this PR as well. This updates the README with general high-level docs of how to use watt as well.

This should inch us closer to a "level playing field" between natively compiled procedural macros and wasm-compiled procedural macros when using a JIT!