[E2E]: Add CAMAsync connector coverage with DeepSeek-V2-Lite

[jiangkuaixue123]

Summary

Add an Ascend NPU end-to-end test for CAMAsyncAFDConnector after the async connector refactor merged in #203.

The test should follow the refactored E2E framework under tests/e2e/ instead of introducing a connector-specific launch or evaluation harness. Use DeepSeek-V2-Lite to keep the model-side resource requirement manageable. The device count is not fixed to three: select and document a representative topology supported by CAMAsync and the available hardware, with more devices allowed when required.

Scope

• Add one stable E2E case, for example async-cam-eager, for CAMAsyncAFDConnector.
• Run on Ascend NPU with DeepSeek-V2-Lite.
• Make the Attention/FFN device lists and topology explicit and configurable through the refactored E2E framework; do not hard-code a 2 Attention + 1 FFN topology for this case.
• If the chosen CAMAsync topology exceeds the current framework's three-device/2A1F assumptions, extend the common topology validation and command construction cleanly rather than adding a connector-specific harness.
• Configure the async connector through the shared runner:
  • connector=CAMAsyncAFDConnector
  • async=true
  • compute_gate_on_attention=true
  • eager execution, since CAMAsync does not support graph execution
• Reuse the shared service startup, readiness checks, GSM8K evaluation, signal handling, and process cleanup in tests/e2e/runner.py and tests/e2e/process_utils.py.
• Add focused CPU-safe unit coverage for the generated commands, connector flags, launch order, configurable topology validation, and failure/cleanup paths.
• Update tests/e2e/README.md and the E2E design/support matrix where needed.

Validation

The PR E2E case should:

• evaluate the standard first 7 GSM8K samples with 8-shot prompting;
• enforce the existing sample-count and accuracy gates;
• fail on missing setup, service failure, incomplete/NaN results, or cleanup failure;
• run without skip or xfail;
• leave no Attention, FFN, or evaluator process behind;
• record the exact Attention/FFN topology, device count, vLLM, vLLM-Ascend, AFD, CANN/CAM operator, model, duration, and accuracy evidence from a real NPU run.

Acceptance criteria

• A stable async-cam-eager pytest node is added using the refactored E2E framework.
• The case uses DeepSeek-V2-Lite.
• The device count and Attention/FFN topology are configurable and reflect the actual CAMAsync validation setup; the case is not restricted to three NPUs.
• Any required relaxation of the common framework's current three-device/2A1F assumptions is implemented and unit-tested in the shared framework.
• The shared runner launches Attention before FFN for CAMAsyncAFDConnector and passes the required async configuration.
• GSM8K validation and strict lifecycle/cleanup checks are reused from the common harness.
• CPU-safe unit tests cover the new case configuration and command construction.
• Real Ascend hardware evidence, including the exact card count and topology, is attached to the implementation PR.
• E2E documentation is updated with the command, prerequisites, limitations, and expected result.

Dependency

Builds on the async connector refactor merged in #203.