feat: add multi-GPU pipeline parallelism via split-mode config to embed addon

[donriddo]

🎯 What problem does this PR solve?

The embed addon currently pins model inference to a single GPU device. On multi-GPU systems this leaves additional GPUs idle, preventing users from leveraging pipeline (layer) or tensor (row) parallelism for embedding workloads.

📝 How does it solve it?

• Adds split-mode config option (none | layer | row) to BertModel setupParams, accepting both hyphen and underscore variants and case-insensitive values. When set to layer or row, the --device flag is omitted so llama.cpp distributes embedding layers across all available GPUs.
• Adds tensor-split config key (comma-separated proportions) for controlling per-GPU layer allocation in split mode.
• Updates tryMainGpuFromMap in BackendSelection to accept both main-gpu and main_gpu variants with conflict detection, matching the LLM addon.
• Adds getEffectiveGpuDeviceCount() to BackendSelection for GPU inventory queries.
• Adds BertModel::getCommonParams() accessor to expose parsed params for testing.
• Adds split-mode and tensor-split fields to GGMLConfig in index.d.ts.
• Bumps version to 0.15.0.

🧪 How was it tested?

• C++ unit tests (test_bert_model.cpp): 8 new split-mode cases covering none, layer, row, case-insensitive input, underscore variant, CPU fallback clears GPU params, invalid value rejection, and both keys reject.
• C++ unit tests (test_backend_selection.cpp): 9 new getEffectiveGpuDeviceCount cases covering zero devices, CPU-only, single dGPU, single iGPU, two dGPUs, dGPU+iGPU, two dGPUs+iGPU, two iGPUs, and accel/CPU ignored.
• JS integration tests (multi-gpu.test.js): 4 tests gated on QVAC_HAS_MULTI_GPU=1 exercising split-mode=layer, split-mode=row, default single-device, and split-mode=layer with tensor-split and main-gpu.

🔌 API Changes

// GGMLBertArgs.config now accepts:
{
  'split-mode': 'none' | 'layer' | 'row',    // optional, default none
  'tensor-split': '1,1',                     // optional, proportional split across GPUs
  'main-gpu': 0 | 'integrated' | 'dedicated' // optional, primary GPU selection
}

[gianni-cor]

Thanks for porting this over from the LLM addon. The core C++ changes look aligned with the upstream LLM implementation, but one important test-wiring difference needs fixing before merge.

packages/qvac-lib-infer-llamacpp-embed/test/integration/multi-gpu.test.js is added, but the embed package's test:integration script still runs only bare test/integration/addon.test.js. The embed PR workflow calls that script directly, so the new multi-GPU integration coverage will not run in desktop CI. In the LLM addon upstream change, test:integration generates/runs the full test/integration/*.test.js set, which is why its new multi-gpu.test.js is exercised.

Please update the embed integration test command/generation path so multi-gpu.test.js is included in CI, or otherwise wire this file into the existing integration test runner.

[gianni-cor]

Please also fix the failing/blocked cpp-lint check before this is re-reviewed.

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (2/1)
- 1 Team Lead OR Management approval ✅ (2/1)

**Bypass rule:** Triggered (2+ Team Lead approvals (Tier 1 exception)). This PR is approved regardless of tier.

---
*This comment is automatically updated when reviews change.*

[gianni-cor]

/review

[gianni-cor]

/review