OpenVINO: PRD-compliant device enumeration and memory reporting for --list-devices

[haarika-madaka]

Summary

This PR implements PRD-compliant OpenVINO device enumeration and memory reporting for --list-devices, scoped to ggml/src/ggml-openvino/ only.

What changed

1. Device enumeration and registration

• Registers one GGML backend device per physical OpenVINO device returned by get_available_devices().
• Filters virtual routing devices by prefix: AUTO, MULTI, HETERO.
• Sorts device identifiers lexicographically for deterministic output.
• Uses exact OpenVINO identifiers as backend device names (no alias synthesis, no normalization).

2. Device selection behavior

• Preserves existing selected device behavior (GGML_OPENVINO_DEVICE with fallback).
• Keeps runtime log line:
  • OpenVINO: using device <...>
• Ensures execution remains bound to the selected OpenVINO device by returning false in supports_op for non-selected registered devices.

3. Device descriptions

• Uses FULL_DEVICE_NAME as device description.
• For NPU, appends architecture when available:
  • <FULL_DEVICE_NAME> (NPU <DEVICE_ARCHITECTURE>)
• Falls back to device identifier on property query failure.

4. Memory reporting (PRD-compliant)

• CPU:
  • Uses existing system RAM path (current behavior).
• GPU:
  • Total = GPU_DEVICE_TOTAL_MEM_SIZE
  • Free = total - sum(GPU_MEMORY_STATISTICS allocations)
• NPU:
  • Total = NPU_DEVICE_TOTAL_MEM_SIZE
  • Free = total - NPU_DEVICE_ALLOC_MEM_SIZE
• Robust fallback:
  • If required properties are unavailable, reports 0 MiB total and 0 MiB free for that device.
• Does not use GPU_DEVICE_MAX_ALLOC_MEM_SIZE.

Scope

Changes are limited to:

• ggml/src/ggml-openvino/ggml-openvino.cpp
• ggml/src/ggml-openvino/ggml-openvino-extra.cpp
• ggml/src/ggml-openvino/ggml-openvino-extra.h

No changes outside OpenVINO backend files.

Validation

Built and ran:

cmake --build build/ReleaseOV --parallel
./build/ReleaseOV/bin/llama-bench --list_devices

Observed output format:

OpenVINO: using device CPU
Available devices:
  CPU: Intel(R) Core(TM) Ultra 7 258V (31623 MiB, 31623 MiB free)
  GPU: Intel(R) Arc(TM) Graphics (iGPU) (29276 MiB, 29276 MiB free)
  NPU: Intel(R) AI Boost (NPU 4000) (31623 MiB, 31623 MiB free)

Notes

• Device identity is strictly the OpenVINO identifier string.
• Registration order is deterministic and does not affect device selection semantics.
• Multi-device execution is not introduced; registration is for discovery/listing.