fix(http): preserve custom finch in sync requests

[mikehostetler]

Summary

• preserve an existing Req.Request finch option instead of overwriting it during non-streaming provider attach
• apply the shared finch preservation logic to explicit attach implementations that bypass ReqLLM.Provider.Defaults.default_attach/4
• add regression coverage for the shared defaults path and sync OpenAI/Azure request preparation

Closes #616

Validation

• mix test test/providers/openai_test.exs test/provider/azure/azure_test.exs test/req_llm/step/retry_test.exs
• mix quality