Consider updating hasher domain separation

[AaronFeickert]

Hasher domain separation is currently handled in a bespoke and fairly complex manner. While the current approach means that domain separation is effectively in a human-readable form, it requires special handling and could be subject to edge cases. It may be a good idea to simplify and harden how this is done.

One approach that is likely to be simpler and less prone to error or edge cases is to ensure that all input entering an underlying hash function be prepended with a type flag that indicates the nature of the input: domain separator, version, label, or arbitrary data. Any type corresponding to variable-length input then includes a little-endian 64-bit encoding of the length, followed by the data itself. Any type corresponding to fixed-length input (like the version, which is always a single byte) does not need a length encoding.

This design is also more flexible, since it supports the addition of new input types via a change to the flag list and even supports things like composition via addition of multiple domain separators or labels.