fix(mnnchat): reuse loaded runtime session for API start

[Juude]

Summary

• reuse an already loaded runtime session when API service starts for the same model
• avoid releasing and reloading the model when chat already opened the model with app config
• add a focused unit test for the API runtime session start policy

Why

Opening the API service after entering chat could hit a config-domain mismatch:

• chat loads the model via useAppConfig=true
• API startup previously called ensureSession(modelId) with the default base-config path
• the runtime controller treated that as non-reusable and released/rebuilt the same model session

That extra reload at the API startup / first request boundary could cause freezes or crashes for local models such as Qwen3.5-0.8B and 2B.

Verification

• ./gradlew :app:testStandardDebugUnitTest --tests com.alibaba.mnnllm.api.openai.service.ApiRuntimeSessionStartPolicyTest
• Focused device validation on 2112123AC before the adb connection dropped:
  • entered chat for Qwen3.5-0.8B-MNN
  • confirmed dumpapp llm status reported an active loaded session
  • started API service for the same model
  • verified logcat hit Reusing loaded runtime session for API start...
  • verified device-local API request returned HTTP 200

Known blockers

• On clean origin/master, ./tests/smoke/scripts/08_regress_api_dumpapp.sh currently fails at baseline because it sources tests/smoke/scripts/smoke_state_helpers.sh, which is not present on that base.
• ./tests/smoke/scripts/09_regress_api_uiautomator.sh rerun on this clean PR branch was blocked after the adb device disconnected from the host.