Put GPT-5.6 Luna first in fast model fallback routing

Problem

The fast-tier fallback chain needed to prefer GPT-5.6 Luna over the previous Claude Haiku-first route. The broader model catalog and routing documentation also needed to remain consistent across the SDK, CLI, templates, skills, and docs.

Implemented

PR #1750 aligns the model catalog, tier inference, fallback routing, generated templates, and documentation:

Fast tier: gpt-5.6-luna -> claude-haiku-4.5 -> gpt-5.4-mini -> gpt-5-mini -> (omit model param).
Premium tier: gpt-5.6-sol first, followed by the supported premium fallbacks.
Standard tier: gpt-5.6-terra first, followed by the supported standard fallbacks.
Replaced retired gemini-2.5-pro with gemini-3.1-pro.
Synchronized the 17-model catalog, runtime constants, inference rules, skills, package templates, CLI examples, and user-facing docs.
Corrected the canonical model-selection source for visual/design route.
Added patch changesets for @bradygaster/squad-sdk and @bradygaster/squad-cli.