fix(native): test b9016 runtime and expose mainGpu

[leehack]

Summary

• Pins the native runtime hook from b8955 to b9016, consuming the published llamadart-native release assets.
• Adds ModelParams.mainGpu and passes it through to llama_model_params.main_gpu for explicit primary GPU selection (Expose main_gpu on ModelParams for explicit Vulkan / CUDA device selection #112).
• Adds ModelParams.splitMode and passes it through to llama_model_params.split_mode, matching upstream llama.cpp semantics for single-GPU selection via ModelSplitMode.none.
• Resolves ggml backend registry/device APIs from the loaded ggml runtime DLL on Windows split bundles when the generated default FFI asset cannot see those symbols.
• Documents the upstream ggml-vulkan cooperative-matrix workaround for driver crashes (GGML_VK_DISABLE_COOPMAT, GGML_VK_DISABLE_COOPMAT2).
• Updates README, website runtime-parameter docs, support matrix docs, and CHANGELOG entries for the b9016/runtime-parameter/loader changes.

Validation

• dart analyze
• dart test test/unit/core/models/inference/model_params_test.dart
• dart test -p vm test/unit/backends/llama_cpp/llama_cpp_service_test.dart
• ./tool/docs/build_site.sh
• GitHub CI passed on docs-only head 577d10a2b.

Issue status

• Blackwell (RTX 50-series, sm_120) unsupported on Windows: Vulkan crash + CUDA runtime too old #111 CUDA / Blackwell: reporter confirmed b9016 works on RTX 5050 Laptop with CUDA.
• Blackwell (RTX 50-series, sm_120) unsupported on Windows: Vulkan crash + CUDA runtime too old #111 Vulkan / Windows split bundle: reporter confirmed the loader fix works on Windows. Diagnostics changed to registeredBackends=[CPU, Vulkan], concrete Vulkan devices, and registryApisUnavailable=false.
• Blackwell (RTX 50-series, sm_120) unsupported on Windows: Vulkan crash + CUDA runtime too old #111 Vulkan / driver crashes: remaining AMD Radeon 860M and NVIDIA RTX 5050 Laptop Vulkan crashes are in vendor cooperative-matrix property queries. Reporter confirmed GGML_VK_DISABLE_COOPMAT plus GGML_VK_DISABLE_COOPMAT2 works around them; this workaround is now documented.
• Expose main_gpu on ModelParams for explicit Vulkan / CUDA device selection #112 device-selection API: addressed at the Dart API/native wiring level with mainGpu plus splitMode.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 29.20354% with 80 lines in your changes missing coverage. Please review.
✅ Project coverage is 76.57%. Comparing base (3b8e101) to head (577d10a).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
lib/src/backends/llama_cpp/llama_cpp_service.dart	27.92%	80 Missing ⚠️

❌ Your patch status has failed because the patch coverage (29.20%) is below the target coverage (70.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #113      +/-   ##
==========================================
- Coverage   77.12%   76.57%   -0.55%     
==========================================
  Files          68       68              
  Lines        8594     8670      +76     
==========================================
+ Hits         6628     6639      +11     
- Misses       1966     2031      +65     

Flag	Coverage Δ
unittests	76.57% <29.20%> (-0.55%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.