Text stream support

[snwfdhmp]

Most models integrate a stream mode to stream text as it is generated.

It is a must-have for llm since it avoids having to wait for the whole output to be finished, which is an issue for long outputs.

How to integrate that into llm ?