fix(perf): support composite (dual-encoder) models in winml perf

[xieofxie]

Problem

winml perf crashed on composite (dual-encoder) models such as SigLIP/CLIP:

winml perf --ep openvino --device cpu -m google/siglip-base-patch16-224 \
  --task zero-shot-image-classification
...
AttributeError: 'WinMLModelForZeroShotImageClassification' object has no attribute 'io_config'

PerfBenchmark assumed every model exposes a single io_config / _session. Composite
models have neither — they orchestrate multiple sub-models (e.g. an image encoder and a
text encoder), each with its own ONNX session. The failure is device-independent: the
(model_type, task) registry routes SigLIP to the composite class regardless of --device.

Fix

Make PerfBenchmark composite-aware while leaving the single-session path's measurement
semantics untouched:

• _aggregate_io_config() — unions the sub-models' inputs (deduped by name, order
  preserved). Their union is exactly the composite forward() kwargs, so random-input
  generation and the info display work unchanged.
• End-to-end timing — composites time the full forward() pass (both encoders + the
  similarity step) via an external PerfStats. Single-session models keep recording
  pure-ORT time inside session.perf(). The monitored loop now takes a run-iteration
  callable so both paths share it.
• Device / EP / task resolved from a representative sub-model.
• _probe_composite_outputs() — runs one forward() and introspects the result so the
  reported outputs are the composite's real task-level tensors (e.g. logits_per_image)
  instead of a deduped union of sub-model ONNX outputs. Best-effort: falls back to the
  aggregated view if the probe fails.

The output describer (_describe_outputs) is architecture-agnostic (handles HF
ModelOutput / dict / sequence / single tensor) — no model-specific field names.

Result

Device:      cpu / OpenVINOExecutionProvider
Task:        zero-shot-image-classification
Inputs:      pixel_values   [1, 3, 224, 224]   float32
             input_ids      [1, 64]            int32
Outputs:     logits_per_image     [1, 1]
             logits_per_text      [1, 1]
             text_embeds          [1, 768]
             image_embeds         [1, 768]

Tests

tests/unit/commands/test_perf_composite.py (new, 15 cases) covers io_config aggregation,
the output describer/probe, input generation, device/EP/task resolution, and the
full-forward() timing path. Existing test_perf_cli.py / test_perf_module.py (31 cases)
still pass — no regression.

🤖 Generated with Claude Code