fix(sandbox): summarize cleanup_all failures with per-agent error reporting

[UGVicV]

Description

This PR resolves the issue where AgentSandbox.cleanup_all() silently failed without reporting which sandbox cleanup failed. Operators could not tell which temporary workspaces remained after bulk cleanup.

Fix

• Changed cleanup_all() return type from None to Dict[str, List] with succeeded and failed keys.
• Each failed entry includes agent_id and error message for operator diagnostics.
• Individual sandbox destroy failures are caught and collected — one failure does not prevent cleanup of other sandboxes.
• Added logger.warning() when failures occur, logger.error() for individual failures.
• Replaced print() with logger.error() throughout the file.
• Removed unused os import.
• Fixed all PEP8 violations (all lines under 79 characters).

Verification (Proof)

Added 3 regression tests in new TestCleanupAll class:

• test_cleanup_all_success: Verifies all sandboxes cleaned up, returned in succeeded list.
• test_cleanup_all_partial_fail: Verifies failed sandbox (mocked OSError) is collected in failed list.
• test_cleanup_all_empty: Verifies empty sandbox returns empty result.

72 passed in 0.07s

• flake8 src/agent/sandbox.py tests/test_sandbox.py -> Passed (0 errors)
• git diff --check -> Passed (0 errors)

Closes #4066.