refactor(auto_agent): extract server schemas, isolate mock server, and add fake evaluator

[ligh-svg]

Part 2/3 for the beam search prototype: refactoring the compilation/profiling backend utilities to separate mock execution on CPU from the production path that runs on TPU VMs. Introducing the FakeKernelEvaluator to simulate compile timings on CPU hosts.

1. common_models.py: Moved shared Pydantic request/response models and utility functions by cpu and tpu server into a common file.
2. mock_tpu_server.py: Added mock_tpu_server to simulate /unified_test, /profile & /autotune on a local CPU host.
3. fake_kernel_evaluator.py: Use LLM to simulates compiler latency and HBM/Compute utilization metrics without TPU.
4. cpu_server.py: clean up to use common_models.py .
5. setup.sh: check the MOCK_COMPILER=true environment variable to determine whether to start mock_tpu_server.py or the real tpu_server.py .