Extend fake motor watchdog and multi-axis fault tests

[s2015-turtle]

Summary

• Keep injected measured state independent of command targets, including zero-gain startup outside operational position limits.
• Model configured CAN watchdog expiry using a monotonic clock; distinguish addressed motion/enable from read traffic and other motor IDs.
• Exercise two IDs sharing vcan, isolated feedback/command loss, per-motor Reset recovery, both existing whole-driver timeout policies, and communication resumption.
• Use bounded readiness/condition waits and RAII control-thread cleanup, replacing arbitrary startup sleeps.

Validation

• Independent source review completed on identical tree fc54c692259ed4a68a30cad1a688fe87174c51b4; no blocking findings.
• Strict C++17 test-support compilation, declaration-stub integration syntax check, and in-memory fake harness passed locally.
• Watchdog deadline ranges do not overlap (200ms/1000ms with 500ms scheduling allowance); a hardcoded 400ms mutant fails the local harness.
• Five ROS distribution build/tests (Humble, Jazzy, Kilted, Lyrical, Rolling) and actual ros2_socketcan/vcan integration passed on latest-main head b156504 (run 36922606740).
• Conflict-free latest-main integration independently reviewed; exact remote tree verified as 7b4bc897e2f870809080ed60a1ae25d71fb7ba4c.

Scope

Test support and documentation only. Existing recovery policy (#40) is unchanged. Direct-driver resumption is not a controller-manager lifecycle reactivation test, firmware conformance, or evidence of physical motor safety; hardware-in-the-loop validation remains separate.

Closes #34