fix(engine): support validated TP x PP topologies

[ShuhaoZhangTony]

Summary

• replace the incorrect TP == device count launcher gate with TP x PP == device count
• validate positive TP/PP sizes and reject conflicting pipeline CLI configuration
• inject --pipeline-parallel-size consistently and record the managed topology

The previous gate made a valid four-card PP2 x TP2 Pipeline Microbatch qualification impossible. This change does not introduce TP1/eager/legacy fallback.

Tests

• bash -n tools/run_vllm_engine.sh: passed
• python -m pytest -q tests/test_systemd_service_scripts.py: 16 passed
• Ruff and git diff --check: passed
• Qwen3.8-27B real NPU0-3 PP2 x TP2 FULL_DECODE_ONLY graph qualification completed through the corrected topology path

Qualification boundary

The Mod was runtime-effective and functionally correct, but matched performance regressed; the tested configuration remains unsupported pending fresh calibration.

Contribution disclosure

AI assistance was used to diagnose the topology gate, edit the launcher, and run tests. Human line-by-line review is required before merge.

[ShuhaoZhangTony]

CI triage: the scoped launcher checks are green (lint, install-linux-cpu, validate-linux-ascend) and the local launcher test file passes 16/16. The full test failure is in four unrelated agentic-workflow knowledge/recommended-reading assertions; frontend is one pre-existing dark-theme screenshot drift. The current main branch is already red (for example run 33836060735). No application/frontend code is changed by this PR, so those baseline failures are intentionally not patched here.