fix: drain native stderr pipe concurrently to unblock EP compilation

[KayMKM]

Why

capture_native_stderr redirected fd 2 into an OS pipe but only drained it
after the wrapped block returned. A native EP that wrote more than the pipe
buffer holds (~64 KB on Windows) blocked in write() and never returned, so the
wrapped block never completed and the drain never ran.

VitisAI model compilation emits several hundred lines of compiler output on
native stderr and reliably filled the buffer, stalling session creation
indefinitely.

Measured on an AMD Ryzen AI 9 HX 370 (VitisAI EP 1.8.72.0, EP cache emptied
before each run):

Scenario	Result
winml perf ... --ep vitisai --precision fp16	ran 17m52s without completing; no compile artifacts written for the last 14m45s
Same model + EP, session created outside the context manager	completed in 185s
Same model + EP, session created inside capture_native_stderr	stalled at the same compilation stage

The last variant isolates the context manager as the cause: same model, same EP,
same session options, same empty cache.

What changed

• capture_native_stderr drains concurrently. A daemon reader thread starts
  before the yield, so the pipe is emptied while the wrapped block runs and the
  buffer never fills. Restoring fd 2 drops the last write-end reference, which
  signals EOF to the reader; it is then joined and the captured lines are
  re-emitted through logging exactly as before. No change to the public
  behaviour (capture, ANSI stripping, [ORT] prefix).

• The eval harness now clears the VitisAI EP compilation cache. It lives
  outside the user profile (C:\temp\<user>\vaip\.cache), so clearing only the
  HuggingFace and WinML caches left compiled NPU artifacts behind. AMD documents
  that these directories must not be reused across VitisAI EP or NPU driver
  versions, and a stale entry also lets a run load a previously compiled model
  instead of exercising the compile path under test.

• Regression coverage. test_no_deadlock_on_large_output writes ~720 KB
  inside the context manager under a 30s timeout, so a reintroduced deadlock
  fails instead of hanging. TestClearDiskCaches covers the new cache directory
  and the None case on non-Windows.

Validation

• pytest tests/unit/utils/test_native_stderr.py — 8 passed, 1 skipped
• pytest tests/unit/eval/test_run_eval_script.py — 111 passed
• ruff check clean on all touched files
• winml perf -m facebook/convnext-tiny-224 --device npu --ep vitisai --precision fp16
  completes in 3m23s (was: no completion after 17m52s)
• run_eval.py --hf-model facebook/convnext-tiny-224 --device npu --ep vitisai
  reports [PASS] 172.4s, perf pass 1/1 (was: [FAIL (TIMEOUT)] 601.9s)