test: cover batched think-tool recovery slot

[jtstothard]

Summary

Add regression coverage for the batched think-tool recovery slot fix from 0a7ad77.

The existing recovery test created a server_slot but left srv.batched_mode false, so it only exercised the legacy non-batched path. This test:

• creates two sessions and server slots;
• enables batched mode;
• starts the real decode_worker_main coordinator;
• targets the second slot explicitly;
• asserts the guard slot is unchanged and only the target slot receives recovery tokens;
• shuts down and joins the coordinator cleanly.

Verification

• git diff --check passed.
• CPU-side compilation of tests/ds4_test.c and dependent objects passed on the Strix host.
• Full make ds4_test is currently blocked because /usr/local/cuda/bin/nvcc is not installed on that host.

Please convert from draft after CUDA-enabled CI or maintainer-side build verification.