fix(cartesia-sdk): avoid retrying wrapped agent

[GautamSharma99]

Summary

The Cartesia wrapper previously enclosed memory preparation and the entire
wrapped-agent stream in one broad try/except.

If the wrapped agent failed—even after yielding output—the exception handler
invoked it again from the beginning. This could duplicate spoken output, tool
calls, messages, payments, or other non-idempotent actions.

This PR:

• extracts best-effort user-turn preparation from agent execution
• limits fallback handling to memory enrichment and persistence preparation
• invokes the wrapped agent exactly once per event
• allows wrapped-agent exceptions to propagate normally
• preserves graceful fallback when memory preparation fails

Behavioral change

Before:

1. The wrapped agent could yield output or perform side effects.
2. The wrapped agent raised an exception.
3. The wrapper caught the exception and restarted the agent.

After:

1. Memory preparation runs with graceful fallback.
2. The wrapped agent runs exactly once.
3. Any wrapped-agent failure propagates without replaying the turn.

Tests

Added regression coverage for:

• failure after partial output
• failure before the first output
• memory-preparation failure falling back to one agent call

Validation completed:

• Cartesia package tests with real dependencies: 4 passing
• isort: passing
• mypy: passing after suppressing Cartesia Line’s missing py.typed marker
• new test files pass Black formatting

The existing agent.py file is not fully Black-formatted on main; this PR
avoids unrelated whole-file formatting changes. The repository-wide Bun suite
also retains unrelated existing failures.