fix(perf): generalize op-tracing availability check and align trace output path

[xieofxie]

Summary

Tightens and generalizes the --op-tracing path in winml perf, plus a QNN profiler session fix.

• Rename is_qnn_profiling_available() → is_profiling_available(resolved_ep, resolved_device, op_tracing). Instead of only probing ORT for the QNN provider, it now validates the benchmark's resolved EP/device/level against the single combination op-tracing currently supports — QNN EP + NPU + basic. EP aliases (e.g. qnn) are normalized, and device/level comparisons are case-insensitive. Supported values live in module constants so they are easy to expand as more tracers land.
• Expose resolved_ep / resolved_device properties on PerfBenchmark so the CLI can gate op-tracing on what actually drove the build/inference, with an error message that reports the resolved values.
• Align the op-trace output path with the benchmark report. The trace JSON now mirrors output: a/b.json → a/b_op_trace.json, instead of being rebuilt from the model name.
• QNN profiler session fix. Build the QNN session via add_ep_for_device(...) for the NPU device (failing fast if it cannot be added) rather than passing providers=[...] directly, and switch provider options to a single dict.

Tests

• Rewrote tests/unit/optracing/test_detection.py for the new signature: supported combo, alias normalization, case-insensitivity, and a parametrized table of unsupported combinations.
• uv run pytest tests/unit/optracing/test_detection.py — 11 passed.