Add encoder that works better for large trees #8

The default encoder has to deeply traverse every object/ array sent over the wire to replace any functions. Another strategy would be to encode the entire object to a binary representation (since that will give us a chance to convert functions into a meaningful representation at the same time), which should also be faster to transfer via postmessage when the messages are very large.

I have an initial branch that does this here: https://github.com/Shopify/remote-ui/compare/cbor-encoding?expand=1. It is mostly inspired by CBOR, borrowing heavily from a ~1.5K JS implementation I found. It also takes some queued from worker-dom’s serialization logic.

Before shipping that branch, we need to:

1. Check its impact on bundle size, and offer some way (probably a Babel plugin) of not paying the cost for both the default encoder and the CBOR encoder
2. See a meaningful performance difference, probably with a puppeteer-based benchmark on a very large tree of remote components