feat: redesign the Reply Drafter prompt for context-aware suggestions

[enaboapps]

Summary

The Reply Drafter prompt asked the model for "3 short, natural, friendly
replies" — a generic instruction that forced every suggestion into one tone.
The feature lands in all kinds of threads (coordinating plans, disagreement,
emotional support, formal/work, declining something), and three cheerful
options are unusable in most of them. For a switch user, unusable suggestions
mean falling back to slow manual typing — the exact thing the feature exists to
avoid.

This replaces LlmManager.PROMPT with a structured 8-step rubric.

What changes

• Varied by stance — the model returns a deliberately varied set, not three
  of the same.
• Always anchored — at least one option always lets the user decline,
  disagree, set a boundary, or buy time.
• Tone- and warmth-matched — replies match the conversation's seriousness,
  and its texting style (kisses like x/xx, emoji, pet names) so they don't read
  as cold.
• 3-5 replies, mixed length — short by default, a longer one when needed;
  "3-5" now agrees with the parser's MAX_SUGGESTIONS = 5.
• The prompt now also tells the model which side of the conversation is the
  user's, and forbids multi-line replies (which the line-based parser would
  otherwise split into bogus suggestions).

Single change — the PROMPT constant. parseSuggestions() and
MAX_SUGGESTIONS are unchanged.

Test plan

• ./gradlew compileDebugKotlin — passes.
• On-device tuning pass: trigger Reply Drafter on casual, tense, formal,
  and affectionate threads — suggestions vary in stance, always include a
  decline/disagree option, match tone, and stay one per line.

Closes #2265