fix: prevent model mixing between LiteLLM configuration profiles

[roomote]

This PR attempts to address Issue #9169. Feedback and guidance are welcome.

Problem

When two configuration profiles are created with LiteLLM API provider but different instances (different base URLs), the list of models was always pulled from one of them, causing issues when they do not provide the same models.

Root Cause

The model cache was using only the provider name as the cache key, not distinguishing between different LiteLLM instances with different base URLs.

Solution

• Added getCacheKey() function that generates unique cache keys including a hash of the base URL for multi-instance providers (LiteLLM, Ollama, LM Studio, Requesty, DeepInfra, Roo)
• Updated writeModels()/readModels() to use unique filenames per instance
• Updated getModels()/refreshModels() to use unique cache keys
• Updated flushModels() to clear both memory and file cache for the specific instance
• Added deleteModelsCacheFile() for cleaning up disk cache

Testing

• All existing tests pass
• Added new tests specifically for multi-instance caching behavior:
  • getCacheKey returns correct keys for different provider configurations
  • Different LiteLLM instances use separate cache entries
  • In-flight request tracking works per-instance

Changes from Previous PR #9170

This PR rescopes the fix as requested, with the following improvements:

• Cleaner implementation with separate getCacheKey() and getCacheFilename() helpers
• File cache is also deleted during flushModels() (not just memory cache)
• More comprehensive test coverage

Fixes #9169

[roomote]

Rooviewer   See task on Roo Cloud

Reviewed the fix for model mixing between LiteLLM configuration profiles. The implementation correctly addresses the root cause by using unique cache keys that include a hash of the base URL for multi-instance providers. Found one minor code quality issue:

• Remove unused readModels() function - The function is defined but never called; disk cache reading is done synchronously in getModelsFromCacheWithKey() instead.

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}

[roomote]

This readModels() function is defined but never called anywhere. The disk cache reading is done synchronously in getModelsFromCacheWithKey() using fsSync.readFileSync() (lines 406-408). Consider removing this dead code, or if it's intended for future async disk cache reading, add a comment explaining the planned use case.

_{Fix it with Roo Code or mention @roomote and request a fix.}