Add --memory flag to winml perf for per-phase memory measurement

[DingmaomaoBJTU]

Summary

Adds process and device memory measurement to winml perf, enabled by default via --memory/--no-memory.

Motivation

Users need to know: "Can this model fit on my device?" This requires two numbers:

• Process peak memory — how much physical RAM the whole pipeline consumes
• Device memory — how much NPU/GPU dedicated VRAM is allocated

Output

Memory:      425.7 MB (process peak) | 54.2 MB (device)

One line, two numbers — everything needed to decide if a model can be deployed. Full per-phase breakdown is still available in the JSON output for advanced analysis.

Implementation

• session/monitor/memory_tracker.py (new): Pure ctypes K32GetProcessMemoryInfo for process memory + single-shot PDH query for device memory. Zero new dependencies.
• commands/perf.py: Integrates MemoryTracker into PerfBenchmark.run() with snapshots at phase boundaries (baseline → load → compile → inference). Extends BenchmarkResult with memory_profile field.
• --memory (default: True): Snapshots are taken between phases, not during the iteration loop, so there is zero overhead on latency measurements. Disable with --no-memory.

JSON output

{
  "memory": {
    "baseline": {"working_set_mb": 125.3, ...},
    "post_load": {"working_set_mb": 389.1, ...},
    "post_compile": {"working_set_mb": 405.8, ..., "device_local_mb": 54.2},
    "post_inference": {"working_set_mb": 412.4, ..., "device_local_mb": 54.2},
    "peak_working_set_mb": 425.7,
    "peak_device_local_mb": 54.2,
    "total_delta_working_set_mb": 287.1
  }
}

Testing

• 15 new unit tests for memory_tracker.py
• All 24 existing perf CLI tests pass unchanged

[xieofxie]

I think we could just enabled it, it doesn't take much time

[xieofxie]

Also consider adding this to --monitor