feat: add per-tool-call timeout to Tool trait

[chaliy]

Summary

The `Tool` trait's `execute()` method has no per-call timeout parameter. If an agent sends a script that runs too long, there's no way for the orchestrating system to abort it except by dropping the future.

Current state

• Resource limits (max_commands, max_iterations) are set on the `BashTool` instance at build time
• `timeout` builtin exists for wrapping individual commands
• But no way for the tool caller to say "this entire call must complete in N seconds"

Proposed API

```rust
pub struct ToolRequest {
pub commands: String,
pub timeout_ms: Option, // NEW: per-call timeout
}
```

Or alternatively:
```rust
async fn execute_with_timeout(
&mut self,
req: ToolRequest,
timeout: Duration,
) -> ToolResponse;
```

Use cases

1. Agent frameworks with per-tool timeouts (LangChain, CrewAI)
2. Preventing runaway scripts in production
3. Different timeouts for different task types (quick query vs. data processing)

Acceptance criteria

• Per-call timeout on Tool execute
• Proper cleanup on timeout (VFS state consistent)
• Clear error message when timed out
• Update tool contract spec