Update on "update llama runner to decode single token"

Right now, we don't print the generated response in the eager runner until all tokens are generated. This is not good experience as we need to wait until all tokens are generated to see the response.

This PR updates it to decode each new token immediately after it is generated.

Differential Revision: [D65578306](https://our.internmc.facebook.com/intern/diff/D65578306/)

[ghstack-poisoned]

examples/models/llama/runner/generation.py

@@ -125,7 +125,7 @@ def text_completion(
             echo (bool, optional): Flag indicating whether to include prompt tokens in the generated output. Defaults to False.
 
         Returns:
-            CompletionPrediction: Completion prediction, which contains the generated text completion.
+            Generated list of tokens.
 
         Note:
             This method generates text completion for the provided prompt, employing nucleus sampling to introduce controlled randomness.