Fix assistant tool-loop resilience

[SebastianBoehler]

Summary

First scoped batch for #43. This PR focuses only on assistant tool-loop reliability and leaves mail/session reuse, model catalog/runtime settings, prompt compaction, and rendering performance for follow-up PRs.

Changes:

• lets cloud assistant calls run bounded multi-round tool loops instead of stopping after one tool round
• returns failed cloud tool calls to the model as tool results instead of aborting the whole turn
• returns failed local tool calls as feedback to the local model instead of aborting the whole turn
• passes fresh memory reads into local tool execution so local and cloud tools use consistent memory freshness
• adds focused tests for multi-round cloud tools, failed cloud tools, and local fresh-memory reads

Scope intentionally left out

• mail IMAP session reuse and partial message fetching
• HTTP client lifecycle reuse/close work
• prompt-size/system-prompt deduplication
• LiteRT timeout/sampling/model download integrity work
• chat rebuild/trace-compaction performance work

Verification

• git diff --check

Not run locally:

• Flutter/Dart formatter, analyzer, and tests. This runtime does not have dart or flutter on PATH.