chore: merge release/v0.3.0 back to main - #1291

…erry-pick to v0.3.0] (#1267)

Cherry-picks #1266 into the v0.3.0 release branch.

This replaces pipe-backed native warning capture with a temporary-file
spool, avoiding EP compiler hangs while preserving warning filtering and
bounded memory use.

Original commit: `e27a0faf` by @KayMKM. Cherry-picked cleanly onto
release tip `e1d6f874` as `ce11819c`.

Release-window changes needed to build a correct v0.3.0 wheel.

### `chore(release): bump version to 0.3.0` (+ uv.lock sync)
Bumps `pyproject.toml` (and `uv.lock`) to `0.3.0`. Version is exposed at
runtime via `importlib.metadata`, so `pyproject.toml` is the only
source.

### `ci(release): pin official build toolchain versions`
`ModelKit Official Build` installed its toolchain with `pip install
--upgrade build twine packaging` — completely unpinned. `build` 1.5.1
(released since v0.2.0) breaks the build inside the OneBranch container:

```
platformdirs\windows.py -> get_win_folder_from_registry
  winreg.QueryValueEx(key, CSIDL_LOCAL_APPDATA)
-> FileNotFoundError: [WinError 2] The system cannot find the file specified
```

The sdist step fails and the wheel step is skipped, so the build
produces nothing. Verified on run
[153739315](https://dev.azure.com/microsoft/windows.ai.toolkit/_build/results?buildId=153739315)
(failed, `build` 1.5.1) vs
[150550609](https://dev.azure.com/microsoft/windows.ai.toolkit/_build/results?buildId=150550609)
(v0.2.0, succeeded, `build` 1.5.0). `build` 1.5.0 and 1.5.1 have
identical `_pip_env()`, so this is environment behavior in 1.5.1, not a
change on our side.

Pinning to the versions that produced v0.2.0 restores a green build and,
more importantly, stops a third-party release from breaking the release
build on publish day.

Validated: official build
[153740145](https://dev.azure.com/microsoft/windows.ai.toolkit/_build/results?buildId=153740145)
succeeded on this branch — sdist, wheel, iKey verification, and PyPI
validation all green.

---------

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

## Summary
- Normalize PathLike model references before telemetry string scrubbing.
- Prevent successful commands using local model paths from exiting with
a Path.replace TypeError during telemetry reporting.
- Add regression coverage for Click options parsed with path_type=Path.
- Cherry-picked as a single focused commit onto release/v0.3.0.

## Validation
- uv run pytest tests/unit/telemetry/test_click_group.py
tests/unit/telemetry/test_utils_scrubbing.py -q (63 passed)
- uv run ruff check src/winml/modelkit/telemetry/utils.py
tests/unit/telemetry/test_click_group.py

## Summary

- Treat TensorRT RTX like OpenVINO when validating Windows PDH GPU
utilization.
- Keep the benchmark, execution provider, device, adapter LUID, latency,
and monitor structure assertions intact.
- Document the unreliable PDH utilization behavior with a TODO in both
ONNX-direct and Hugging Face GPU cases.

## Validation

- `uv run --no-sync pytest -m e2e -q
"tests/e2e/test_perf_e2e.py::TestPerfONNXDirect::test_benchmark_ep_device_gpu[nv_tensorrt_rtx]"
"tests/e2e/test_perf_e2e.py::TestPerfHuggingFace::test_benchmark_ep_gpu[nv_tensorrt_rtx]"
--basetemp=temp/pytest_tmp/release-v030-trtrtx-monitor-fix` (2 passed)
- `uv run --no-sync ruff check --fix tests/e2e/test_perf_e2e.py`

Adds the v0.3.0 CHANGELOG entry, grouped by user-visible behavior
changes, improvements, fixes, internals, and release assets.