fix(server): preserve tool results and reduce checkpoint churn

[ahmedhesham6]

Summary

This PR fixes two server-side issues observed in gateway/session runs:

1. Tool result linkage was lost in session storage conversion

   • Preserves tool_call_id when converting LLMMessage::ToolResult -> ChatMessage.
   • Adds regression coverage in server message storage adapter tests.

2. Checkpoint churn in short sessions

   • Reworks checkpoint runtime persistence to use a dirty + signature-based snapshot strategy.
   • Hooks now update in-memory latest messages; persistence is deduplicated and flushed via snapshot path.
   • Reduces excessive checkpoint creation while preserving terminal/error persistence behavior.

Also added a TODO in session_actor to migrate checkpoint storage to native Vec<stakai::Message> and remove adapter conversion.

Why

• Prevents tool-call/result chain breakage in stored sessions (which can surface as aborted/missing tool response behavior).
• Prevents runaway checkpoint growth in sessions with small message volume.

Validation

• cargo fmt on touched files
• cargo clippy -p stakpak-server -p stakpak-shared --lib -- -D warnings
• cargo test -p stakpak-shared tool_result -- --nocapture
• cargo test -p stakpak-server preserves_tool_result_tool_call_id_through_storage_adapter -- --nocapture
• cargo test -p stakpak-server session_actor::tests -- --nocapture
• cargo check -p stakpak