Learning panel follow-up

[minestarks]

DON'T REVIEW will split into separate PRs.

Addresses remaining feedback from #3228.

Structured hint markers

• Katas exercise hints previously used <details><summary>Need a hint?</summary>...</details> HTML blocks, extracted by a fragile regex in generate_katas_content.js.
• Replaced with a :::hint / ::: fenced block syntax across all 49 exercise index.md files. The generator now matches these structured markers instead.

AI file consolidation

• agents/, prompts/, skills/, and resources/chat-instructions/ were scattered across vscode. Moved all Copilot-related files into a single ai directory.
• Updated package.json contribution paths (chatInstructions, chatSkills, chatPromptFiles, chatAgents) accordingly.
• The QDK Learning agent model was hardcoded to "Claude Haiku 4.5 (copilot)". Now specifies ["Claude Haiku 4.5 (copilot)", "auto"] so VS Code falls back gracefully if the model is unavailable.

Progress file error handling

• loadProgress() previously caught all errors (including corrupt JSON) silently and reset progress to defaults, destroying user data.
• Now distinguishes file-not-found (use defaults) from invalid JSON or unexpected structure (throw with a descriptive message). The corrupt file is preserved — the user must fix or delete it.
• saveProgress() wraps write failures with a clear error message. Errors propagate through existing paths:
  • Panel UI: caught by handleAction → shown in the webview.
  • LLM tools: caught by invoke() → surfaced as CopilotToolError to the model. get-state also returns the error string when initialization fails due to a corrupt file.
  • Dispose: shows a showWarningMessage notification instead of silently swallowing.