Refactor to reuse stream parsing across ChatModels

[jackmpcollins]

Big refactor. Replaces near-duplicate generator functions with better abstractions.

To parse an LLM streamed response now requires

• StreamParser to identify string vs tool and parse these
• StreamState to keep track of current message snapshot, usage, etc. for a streamed response

This removes the assumption that the LLM returns either string or tool calls, in preparation for #232

Improved parsing of output types. So now FunctionCall or ParallelFunctionCall is required in the return type of the prompt-function rather than allowing these to be returned when not in the type.

Also added new error types FunctionCallNotAllowedError, ObjectNotAllowedError, UnknownToolError which can be added to the LLM-assisted retry logic in future.