feat(option): Negative prompt

[grencez]

Reference: https://www.github.com/ggerganov/llama.cpp/pull/2135

• feat(option): Custom sampling order #36 - Wait for sampling refactor.
• Implement.

[grencez]

• Implement as any number of prompts with different weights.
• Different prompts associated with different chat prefixes.
• Different context lengths, so the tokens can be steered more effectively.
• Different histories, potentially making room for private "inner monologue".

I have to think carefully about this one. It seems very tricky to balance multiple contexts.

[grencez]

We could call this an agent. The general outline could be as follows (but needs some work):

(rolling_prompt (())
 (file "rolling.txt")
 (loop (())
  (turn (prefix "{{user}}: ") (user_input +true))
  (turn
   (prefix "least_unhinged_twitch_chatter:")
   (guidance (agent polite_speech) (scale -1.0)))
 )
)
(agents (())
 (polite_speech
  (priming_prompt (())
   (text "You are a kind person named least_unhinged_twitch_chatter who uses speech compatible with a polite society.\n")
   (main_rolling_prompt)  ; Do we need the whole thing?
 ))
)

The sampling customization in #36 may affect this, so it's blocked on the same thing.