[Feature Request: ChatGPTClient] Add option to avoid buildPrompt and use standard messages payload

[danny-avila]

I might work on this myself, but I'm highlighting the discrepancy between building the prompt and making the standard API call with all messages as array in payload

In each test, I use the same set of messages.

ChatGPTClient
initial message prompt tokens: 58
followup prompt tokens: 82
2nd followup prompt tokens: 114
final: 179

Raw API Calls
initial message prompt tokens: 9
followup prompt tokens: 32
2nd followup prompt tokens: 63
final: 127

The results I get are very comparable, and I could save tokens this way.

---

With System Message

Note that the current buildPrompt method is more useful when you want a system message, as it's comparable in token usage, and it often nets better results for gpt-3.5.

ChatGPTClient
initial message prompt tokens: 47
followup prompt tokens: 111
2nd followup prompt tokens: 194
final: 280

Raw API Calls
initial message prompt tokens: 34
followup prompt tokens: 98
2nd followup prompt tokens: 181
final: 267