Add Hrana 3

[honzasp]

This PR introduces and implements version 3 of the Hrana protocol. The changes relative to Hrana 2 are as follows:

• Added Protobuf encoding in addition to the existing JSON encoding. The encoding is negotiated in the same way as the protocol version.
• Added cursors. Previously, results from a statement or a batch of statements had to be loaded into memory and serialized into a single large message. Cursors support incremental reading of results (with row granularity) from executing a batch.
• Added request and batch condition to query the autocommit status. This is necessary to correctly implement transactions in libsql clients, to make sure that if the transaction is implicitly rolled back (for example by using OR ROLLBACK conflict resolution strategy), we won't execute follow up statements in autocommit mode.

These additions stretch the Hrana protocol near its breaking point. At the beginning, Hrana was small and simple protocol, but it has gotten quite bloated: there are two protocols (WebSocket and HTTP), two encodings (JSON and Protobuf) and multiple ways to execute statements (execute request, batch request and cursors). Future versions of Hrana should simplify the protocol.

Fortunately, clients can be simpler, because they need to implement just a single combination of protocol and encoding (typically WebSocket and Protobuf).

---

I also rewrote the Hrana spec, so Hrana 3 is described in a single document (and not as a set of changes to the previous versions, like Hrana 2).

[MarinPostma]

LGTM, few minor nits

[MarinPostma]

You can't assume that auto-commit is true from an invalid state

[honzasp]

I was hoping that the invalid state would not occur in practice :) Assuming autocommit is safer, because it is more dangerous to execute a statement that is intended to be in transaction outside of a transaction. (I.e., it is better if the client aborts a transaction if the server reaches the invalid state.)