feat(hardware): support non-NVIDIA discrete GPUs with DirectML and launch flag injection (M11)

[BerryUIKI]

Summary

This PR implements Milestone M11: Local Inference on Selected Non-NVIDIA Discrete GPUs.

Key Capabilities

• Multi-Vendor Discrete GPU Detection:
  • Unified detection across NVIDIA (CUDA), AMD Radeon (DirectML on Windows, ROCm on Linux), Intel Arc (DirectML on Windows, OneAPI IPEX on Linux), and Apple Silicon (Metal Performance Shaders on macOS).
  • Queries Windows WMI and Linux
    ocm-smi\ without importing heavy dependencies (no PyTorch, no CUDA C-extensions).
• Classification & Reference Hardware Tiers:
  • Distinguishes Tier 1 Verified Reference Hardware (Radeon RX 7900 XTX/XT, RX 7800 XT, RX 6700 XT, Intel Arc A770, Arc A750) from experimental or unverified cards.
  • Automatically recommends Cloud BYOK (Fal.ai, SiliconFlow, OpenAI) when discrete GPU compute is unavailable or insufficient (< 6GB VRAM).
• Dynamic Launch Flag Injection:
  • ComfyUI supervisor automatically passes vendor-optimized launch flags (--directml, --use-split-cross-attention, --lowvram, --force-fp16).
• Support Matrix Documentation:
  • Comprehensive reference guide in \docs/SUPPORT_MATRIX.md.

Automated Verification

• Backend pytest: 99 passed (including 4 new tests in \�ackend/tests/test_m11_hardware.py)
• Frontend: TypeScript typecheck and Vite build passed cleanly
• Launcher: \cargo test\ passed cleanly