docs(prompt-guide): fold customer-validated learnings into prompt optimization guide

Restructures the Vapi Prompt Optimization Guide around production learnings
from docs/learnings/*. Adds platform-aware sections that previously lived
only in the learnings folder: Liquid trust tiers (and static parameters as
the LLM-invisibility primitive), squad re-entry protocol, two-agent
voicemail relay, idle-message vs voicemail-silence deadlock, gpt-5
reasoning-vs-chat selection, Cartesia em-dash/SSML gotchas, verbose-banlist
priming, tool resource ID TTS leak, time discipline via hooks (not prompts),
and a section-by-section cross-reference index.

Author: dhruva-vapi