fix(models): preserve live tools in retry request snapshots

[zakahan]

RetryingLiteLlm can fail before making a model request with TypeError: cannot pickle '_asyncio.Future' object when a tool in LlmRequest.tools_dict holds asynchronous runtime state. This was observed in a Skill sandbox after tool responses were returned, when the agent tried to start its next model turn. The unconditional retry snapshot deep-copies live tool instances even when no HTTP 429 occurs.

Preserve tool instance identity through the deepcopy memo while continuing to deep-copy request contents, configuration, and the tool mapping. This keeps the retry payload isolated from first-attempt mutations without cloning tool sessions or asynchronous state. Retry eligibility and fallback behavior remain unchanged.

Validation:

• Added four regression cases covering streaming/non-streaming requests with and without a pre-output 429. All four reproduce the Future error before the fix and pass afterward.
• Regression checks verify live tool identity and isolation of nested contents, config, and tool-map mutations on retry.
• pytest tests/models -q: 16 passed, including existing no-replay-after-output and non-429 tests.
• Pre-commit on both changed files: Ruff lint/format and Gitleaks passed.

No live model calls were used for these tests.