Fix flaky test_ash_end_to_end

[zigpy-review-bot]

test_ash_end_to_end fails intermittently on CI, most recently on dev in https://github.com/zigpy/bellows/actions/runs/36362650409 (Python 3.14, FakeTransportWithDelays: TimeoutError instead of NotAcked), and earlier in https://github.com/zigpy/bellows/actions/runs/34772581160 and https://github.com/zigpy/bellows/actions/runs/30421043625 (DID NOT RAISE TimeoutError at the "ncp failure" step, fixed by a rerun).

The test scales the ASH ACK timeouts down to 0.4–3.2 ms and used wall-clock sleeps to decide how long a link stays down, so a short CPU stall on a busy runner changes the outcome. This is a test-only change; bellows/ash.py is untouched.

• "host failure" / "ncp failure" steps: the transport was unpaused after a fixed sleep(t_rx_ack * 15). If the last of the 5 attempts had not been sent yet by then, it got through and no TimeoutError was raised. The link now stays paused until the send has actually failed.
• Reset: a fixed sleep(0.01) after sending RST did not guarantee the RSTACK had been processed, so the next send could hit NcpFailure. The test now waits (bounded to 1 s) until the host is connected again.
• Excessive-NAK step: each NAK raced a sub-millisecond host ACK timeout. If the final attempt timed out first, TimeoutError came out instead of NotAcked (the failure in the linked run). The host now gets a 5 s ACK timeout for this step, and FakeTransportRandomLoss drops nothing during it, so every attempt ends in a NAK. The NAK retry and final-NAK paths in _send_data_frame are still covered (ash.py coverage unchanged).

Stress results with pytest-repeat + pytest-xdist on an 8-core machine, test_ash_end_to_end only:

Load	dev	this PR
-n 8, 6000 runs (CI-like)	2 failed	0 failed
-n 24, 1600 runs (3× oversubscribed)	178 failed	0 failed

Under even heavier load (-n 32, 4000 runs) this PR still failed 2 runs. Both were sends that are expected to succeed but ran out of their ~14 ms retry budget. Fixing that would mean raising the patched T_RX_ACK_MAX, which made each parametrization about 5× slower (≈0.3–0.45 s), so it is left out.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.58%. Comparing base (c93fc1b) to head (0b3d19f).

Additional details and impacted files

@@           Coverage Diff           @@
##              dev     #760   +/-   ##
=======================================
  Coverage   99.58%   99.58%           
=======================================
  Files          64       64           
  Lines        4330     4330           
=======================================
  Hits         4312     4312           
  Misses         18       18           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.