bug(core): unbounded _execute recursion on sandbox_expansion_required can loop forever (scheduler.ts)

## What happened?

`packages/core/src/scheduler/scheduler.ts:927-934` recursively calls `this._execute()` when a tool returns `sandbox_expansion_required`, with no counter or depth limit:

```ts
// Call _execute synchronously and properly return its promise to loop internally!
return await this._execute(
{ ...activeCall, status: CoreToolCallStatus.Scheduled } as ScheduledToolCall,
signal,
);
```

A malicious or buggy tool returning `sandbox_expansion_required` on every call recurses forever (stack growth / infinite loop). The surrounding `catch {}` falls through silently.

## What did you expect to happen?

Bounded retry: iterative loop with `maxSandboxExpansions` (e.g. 2–3), then fail the call with an error and surface retry count in telemetry. Abort signal should also break the loop.

## Client information

- Repo `google-gemini/gemini-cli` @ `main` `9c1b0a610`
- Area `packages/core/src/scheduler/scheduler.ts:927-934`
- Platform: source checkout

## Login information

N/A — scheduler logic; repro with a stub tool always returning `sandbox_expansion_required`.

## Anything else we need to know?

Fix direction: replace recursion with loop + counter, emit `sandbox_expansion_retry` telemetry, add regression test asserting 3rd consecutive request fails instead of recursing. Related hardening track #29214/#29216.

Start in packages/core/src/scheduler/scheduler.ts:927-934 and trace the _execute path for sandbox_expansion_required. Reproduce the behavior with a stub tool that always returns that status, then add a regression test showing the third consecutive request fails, the retry count is reported in telemetry, and the abort signal breaks the loop.

Written by the indexing model from the issue text.