feat(rknn): add shared RV1126B hardware acceleration support

[samuel--hu]

What changed

• adds a data-driven Rockchip platform layer for RK3576 and RV1126B instead of maintaining a separate backend per chip
• stages platform-specific RKNN/media runtimes from locked manifests and validates model/toolchain compatibility
• extends the shared RKNN + MPP DMA-BUF + RGA path, including native INT8 bound-input preprocessing and hardware-path counters
• exposes accelerator/platform identity through the existing system API and UI
• adds qualification tooling, ScenarioBench hardware gates, tests, and bilingual documentation

Why

RV1126B models are not binary-compatible with RK3576 models, but the runtime architecture, media pipeline, preprocessing, telemetry, conversion workflow, and product UI can remain shared. This PR isolates only platform facts and artifacts in manifests while keeping one Rockchip implementation.

Validation

• RV1126B exact candidate dbb35cc1: native build, device test binary, package, deployment, login, video/task/event/system customer journey
• RV1126B device tests: 47 cases / 1,900 assertions passed
• clean four-channel run: approximately 7 FPS per channel at target 5 FPS, zero discarded frames, four algorithm preview streams healthy, no new OOM, engine ERROR/FATAL, or CPU fallback markers
• hardware counters observed: RKNN forward, MPP DMA-BUF decoder frames, RGA-bound native INT8 frames, classifier DMA-BUF crops; expected decoder and encoder backend rockchip-mpp-rga
• a prior completed 3-minute four-channel journey passed the strict checkpoint; the final shortened run was manually stopped per request and therefore its strict audit is FAIL only on run.status and final teardown delay, not workload/performance checks
• x86 regression: 933 cases / 200,402 assertions passed (final source differs from the built semantic candidate only by clang-format in one qualification source file)
• RK3576 exact build/test/resource/frontend/package regression passed
• ScenarioBench 75/75; Python workflow suites, docs verification, and frontend build passed
• changed-scope clang-format passed; the full formatter still reports 10 pre-existing baseline files outside this change

Scope boundary

This establishes an RV1126B development candidate and demonstrates adaptation quality. It is not yet a multi-platform production certification or a long-duration stability claim.

Test note

An earlier harness attempt retained three differently named four-channel benchmark sets because cleanup was disabled, leaving 12 configured channels and causing an OOM while opening the third preview. Those generated channels were removed without backup, the engine was clean-restarted, and the four-channel result above was collected from a zero-channel baseline.