Show NPU memory in winml sys

[xieofxie]

Summary

• continue Show dedicated and shared GPU memory in winml sys #1421 by exposing DXCore dedicated and shared memory for NPU adapters in winml sys
• include dedicated_memory_mib and shared_memory_mib in NPU JSON and render both capacities in text output
• preserve the existing GPU WMI fallback when DXCore is unavailable

Validation

• uv run ruff check --fix src\winml\modelkit\commands\sys.py tests\unit\commands\test_sys_device_info.py tests\e2e\test_sys_e2e.py: passed
• uv run pytest tests\unit\commands\test_sys_device_info.py tests\unit\sysinfo\test_dxcore_adapters.py -q: 25 passed on AMD
• uv run pytest -m e2e tests\e2e\test_sys_e2e.py -q: 22 passed on AMD, Intel, and Qualcomm
• all pull-request CI checks passed

Feature results

Feature	Hardware	Observed NPU memory	Feature validation
amd_npu	AMD NPU Compute Accelerator Device	0 MiB dedicated, 15,932 MiB shared	--ep vitisai --device npu perf E2E passed in 120.44s
ov_npu	Intel AI Boost NPU	0 MiB dedicated, 18,376 MiB shared	22 sys E2E tests passed; --ep openvino --device npu perf passed in 51.36s
qnn_npu	Snapdragon X Elite Hexagon NPU	0 MiB dedicated, 7,989 MiB shared	Ruff, 20 sys unit tests, and 22 sys E2E tests passed; --ep qnn --device npu perf passed in 49.08s (0.394 ms mean, 2,541.3 samples/s across five timed inferences)

[Qiong Wu (qiowu) (DingmaomaoBJTU)]

Reviewed the complete diff and surrounding DXCore inventory and device rendering paths. No blocking findings. NPU memory uses the existing optional DXCore capacity fields, zero is rendered correctly, and the GPU WMI fallback remains intact. Verified all current PR checks are successful and the commands CI log confirms all 20 device-info tests passed. Hardware E2E results are reported by the author; not independently rerun locally.