Replace byte[] with ReadOnlyMemory/Span where possible

[lcodes]

For large network payloads (hearing cases of 75k+ rows in a single update) the receiving thread currently breaks down each of these rows into a separate byte[] allocation. This is to replace them with slices into existing buffers.

The tricky part is storage inside caches, where the byte[] of the row is used as a key; a solution is to make table types produce implement hash and equality operators to become their own keys.