Support parallel function invocation

[mbroecheler]

LLMs are starting to support parallel tool invocation where the model responds with multiple functions to execute.
This allows the LLM to pull in more information at once reducing the number of iterations and thereby response latency.