Change last_n_messages to default to 'auto'

[afourney]

Why are these changes needed?

An 'auto' mode was added to the last_n_messages code execution configuration in #693. Since then it has been my top recommended setting for people struggling with GroupChat. Basically, if last_n_messages is set too low (e.g., the current default of 1), then code execution can break if another agent speaks between when code is written and when it is set to be executed. 'Auto' avoid this problem.

In non-GroupChat settings, last_n_message = 'auto' and last_n_message = 1 work identically. There would be no change in behavior.

TL;DR;
Though 'auto' can vastly improve GroupChat performance (e.g., see my experiments with HumanEval), it is still not the default. This PR changes that.

Related issue number

#693

Checks

• I've included any doc changes needed for https://microsoft.github.io/autogen/. See https://microsoft.github.io/autogen/docs/Contribute#documentation to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.

[codecov-commenter]

Codecov Report

Attention: 3 lines in your changes are missing coverage. Please review.

Comparison is base (05923a7) 32.45% compared to head (214012e) 14.18%.

Files	Patch %	Lines
autogen/agentchat/conversable_agent.py	0.00%	3 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1356       +/-   ##
===========================================
- Coverage   32.45%   14.18%   -18.28%     
===========================================
  Files          41       39        -2     
  Lines        4905     4899        -6     
  Branches     1119     1120        +1     
===========================================
- Hits         1592      695      -897     
- Misses       3187     4178      +991     
+ Partials      126       26      -100     

Flag	Coverage Δ
unittests	14.18% <0.00%> (-18.23%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[afourney]

Just to add some data to support this move.

I used AutoGenBench to run a random 10% subsample of HumanEval through the GroupChatThreeAgents_Distractor scenario in two conditions: last_n_messages = 1 (the current default), and last_n_messages = "auto". I used GPT-3.5-turbo-16k. The results are as follows:

                 Trial 0             Trial 1
		 last_n = 1    last_n = auto
Task Id          Success             Success
-------------  ---------           ---------
HumanEval_107      False                True
HumanEval_22       False                True
HumanEval_43        True                True
HumanEval_88        True                True
HumanEval_14       False                True
HumanEval_157      False                True
HumanEval_141      False                True
HumanEval_57       False                True
HumanEval_154       True               False
HumanEval_153      False                True
HumanEval_93       False                True
HumanEval_137      False               False
HumanEval_143       True                True
HumanEval_13        True                True
HumanEval_49        True                True
HumanEval_95       False                True
-------------  ---------           ---------
Successes              6                  14
Failures              10                   2
Missing                0                   0
Total                 16                  16