[NPU] Clean up DSV4 async CAM follow-ups from #264

[jiangkuaixue123]

Background

PR #264 added the DSV4 NPU async CAM/ubatching path. A few tactical workarounds and duplicated flows were intentionally left for follow-up so that their runtime contracts can be validated separately.

Related: #264 and #150.

Follow-up tasks

• Remove the device-wide torch.npu.synchronize() from the connector-driven FFN path. Replace it with the narrow completion/credit mechanism required to keep CAM buffers alive and prevent premature reuse. Coordinate this with the shared stream-pipeline work in [Feature]: Add CAMAsync multi-stream execution on a shared NPU stream pipeline #150.
• Remove the direct register_afd() call from get_afd_model_config(). Model registration should be established through the appropriate plugin or worker initialization path rather than as a side effect of model-config conversion.
• Investigate why the DSV4 async deployment requires spawn. Provide a minimal reproducer and identify which component depends on it. If it is not required, remove AFD_FORCE_SPAWN_MULTIPROCESSING and the process-wide multiprocessing override. If it is required, isolate it to the affected NPU runtime/launcher path and avoid modifying Python's global/private multiprocessing context from package import.
• Low priority: compare the DSV4 async flow with the existing DSV2 async CAM implementation and extract genuinely shared scheduling, metadata, and connector logic while keeping model-specific behavior explicit.

Validation

• Cover buffer lifetime and ordering after removing the synchronization, including multiple layers/stages and repeated in-flight work.
• Verify model resolution in spawned workers after removing the direct register_afd() call.
• Test the supported multiprocessing methods on the affected Python/CANN/CAM environment and document the result.
• Keep existing DSV2 and DSV4 async CAM unit/E2E behavior unchanged after refactoring.

Acceptance criteria

• No device-wide synchronization remains in the DSV4 connector-driven FFN steady-state path unless a demonstrated external CAM contract requires it.
• Model-config conversion no longer invokes register_afd() directly.
• The spawn requirement is either removed or backed by a reproducible, narrowly scoped runtime contract without a package-import-time global multiprocessing monkey patch.
• Any shared DSV2/DSV4 abstraction reduces real duplication and preserves explicit model-specific differences.

Before submitting

• I searched existing issues and RFCs.
• I identified this as plugin-owned NPU model/runner initialization and cleanup work; no vLLM source-tree modification is expected.

[jiangkuaixue123]

cc @bjf-frz @ShwStone