Signed-off-by: Alexey Rusakov <ktirf@users.sf.net>

All the other methods clear their random inputs. This one needs to do the same,
to reduce the risk of the randomness being used elsewhere and leaking key info.

Some clients expect the session id to be globally unique,
so allowing the end devices to pick the session id will cause
problems.

Include the current ratchet index with the initial keys, this decreases
the risk that the client will supply the wrong index causing problems.

Sign the initial keys with the ratchet ed25519 key, this reduces the
risk of a client claiming a session that they didn't create.

Add ed25519 keys to the inbound and outbound sessions, and use them to sign and
verify megolm messages.

We just stuff the ed25519 public key in alongside the megolm session key (and
add a version byte), to save adding more boilerplate to the JS/python/etc
layers.

Putting the session_id inside the packed message body makes it hard to extract
so that we can decide which session to use. We don't think there is any
advantage to having thes sesion_id protected by the HMACs, so we're going to
move it to the JSON framing.

Factor the actual message encoding/decoding and encrypting/decrypting out to
separate functions from the top-level functions which do the base64-wrangling.

This is particularly helpful in the 'outbound' code-path where the offsets
required to allow room to base64-encode make the flow hard to see when it's all inline.

Initialise megolm_cipher via the preprocessor macro, instead of with a
function.

memset is at risk of being optimised away, so use _olm_unset instead.

We need to be able to inspect an outbound session so that we can tell our peer
how to set up an inbound session.