feat: add AI Apps cleanup category for Ollama and LM Studio

[brainx]

Summary

Add a dedicated AI Apps cleanup category for known local AI runtimes on macOS, starting with Ollama and LM Studio.

Why

PureMac already cleans broad cache locations, but local AI apps now generate meaningful amounts of removable app-specific data.

This change makes that cleanup more explicit and safer:

• users can inspect AI-app residue in its own category
• app-specific cleanup paths no longer have to rely on generic cache buckets
• primary user data is kept out of cleanup scope

Scope

Included cleanup targets:

• ~/.ollama/logs
• ~/Library/Caches/ollama
• ~/Library/Caches/com.electron.ollama
• ~/Library/WebKit/com.electron.ollama
• ~/Library/Saved Application State/com.electron.ollama.savedState
• ~/.lmstudio/server-logs

Explicitly excluded from cleanup:

• ~/.ollama/models
• ~/.lmstudio/models
• ~/.lmstudio/conversations

Implementation

• add a new AI Apps category to the cleanup model
• route the category through ScanEngine
• scan conservative Ollama and LM Studio cleanup paths only
• reduce double-counting by excluding AI-app cache roots from the broad User Cache scan
• keep non-AI Google app caches visible by scanning the Google vendor root instead of only Google/Chrome
• update English and Simplified Chinese localization
• update README feature and safety documentation

Clarification on the Google cache change

The Google cache handling is not part of the AI-app cleanup scope itself.

It is a follow-up safety adjustment to the generic User Cache deduplication logic:

• the broad ~/Library/Caches scan now excludes vendor roots that are handled more explicitly elsewhere
• if ~/Library/Caches/Google is excluded from that broad scan, re-adding only Google/Chrome would accidentally hide other Google app caches from scan results
• scanning the Google vendor root keeps unrelated Google app caches visible while still preventing double-counting in the generic cache pass

Validation

• built locally with ./script/build_and_run.sh --verify
• confirmed the project still compiles successfully after the category and scanner changes

Notes

• This PR intentionally does not include local helper files or Codex run-button configuration.
• Manual verification on a machine with Ollama and LM Studio installed would still be useful before marking ready for review.

[momenbasel]

LGTM. Well-scoped AI Apps category — conservative cleanup targets (logs/caches only), explicit model exclusions, deduplication helpers, proper double-count prevention in User Cache. One note: zh-Hant localization (PR #24) will need the new "AI Apps" key added separately.

[momenbasel]

Excellent work @brainx! The deduplication logic, path normalization, and explicit model exclusions show real attention to detail. The CleanupTarget pattern you introduced is already being referenced as a template for future categories. Thanks for the high-quality contribution!

[brainx]

@momenbasel Thanks for feedback and merge 👍