feat: Use FastCDC algorithm for chunking (#29)

To minimize the number of chunks that need to be resent from server to client use content defined chunking.  

Uses a FastCDC implementation based on https://github.com/ronomon/deduplication/blob/master/binding.js

Currently uses an average chunk size of 1KiB, as this is the size of value we recommend to users and perf test.   In practice this achieves a very high percentage of deduplication across document versions (typically 97+% for typing in documents > 100KiB).

Updates the provider to use experimentalWatch instead of subscribe to avoid the need to deserialize unchanged chunks (from base64 strings to Uint8Array).

Also optimizes the client side mutator implementation to just write the client update value and not merge with the server update.  We only apply the client update on initial load of a document as all other client updates will have originated from that ydoc instance and so do not need to be applied.