Make it possible to keep some chats after downgrading GPT4All

[cebtenzzre]

trafficstars

What this PR fixes:

• Deserialization didn't explicitly handle future versions, so opening older GPT4All with newer chats saved was ill-defined (usually hangs).
• Chats were saved unconditionally, so updating GPT4All automatically upgraded all of your chats such that previous versions cannot read them.

Other things I fixed while trying to pin down Q_UNREACHABLE crashes:

• A missing serialization case for API models, which happened to work before because of compiler optimization of Q_UNREACHABLE in Release builds.
  • This would cause Debug builds to always crash on exit if you had ever previously saved an API chat with a Release build.
• GPT-J state from v2.4.x and older was incorrectly recognized as Llama state, among other conflicts.
• m_llModelType was not initialized when a Chat is constructed, which could cause UB if it was ever read.
• If a chat file had an unsupported model type, we would either crash or load the chat anyway.

Out of scope for this PR:

• #2626 (saw while testing)
• Saving chat files eagerly rather than at exit
• Handling SIGINT to save chat files when interrupted from command line
• Not losing the chat if the first message is still generating while we exit