support sequential tool calls and fix mistral

[deathbyknowledge]

• Mistral errors if an assistant message includes both content and tool_calls so this adds a check for it.
• Mistral also errors if a tool call does not include tool_call_id, so this includes it.
• Add a model setting sequentialCalls, so a model is allowed to chain tool calls together, instead of only when the last message is a user message. This makes all responses be a simulated stream.

Usage:

const workersAI = createWorkersAI({ binding: env.AI });
const model = workersAI(Model.Mistral24B, { sequentialCalls: true })

[andyjessop]

Hey @deathbyknowledge, thanks for this! I've got a couple of questions.

• If this is fixing something that breaks normal usage, then is it the right way to go to hide the functionality behind a model setting? I'm worried here that no-one will discover this setting and just assume that the model is broken.

• Would you be able to add some example JSON objects to the description to explain exactly what the issue is here, because I'm interested to see if there's a general solution.

• Is this an issue that is just for Mistral? And if so, is it the job of the workers-ai-provider to fix it, or is it a model output issue?

I think generally I'm missing some context that defines the exact problem that we have here. And it would be good to detail it for posterity.

[threepointone]

revisiting this, I'd like to land support for mistral, can we address feedback and move forward here?