fix: anchor behave step start_time to formatter.step() callback

[gibiw]

Summary

• parse_step computed start_time = current_time - step.duration, where current_time was the moment of behave's formatter.result() callback. That callback fires after the after_step hook, so any hook latency inflated both the recorded start_time and end_time — durations stayed correct but absolute timestamps drifted forward by the cumulative hook overhead. In a scenario with a 200 ms after_step, the first step appeared to start 200 ms into the scenario rather than at the scenario boundary.
• Capture wall-clock in QaseFormatter.step() (behave's "right before this step executes" formatter hook) and pass it explicitly to parse_step via a new start_time argument. Compute end_time = start_time + step.duration so the window matches the real step body and excludes after_step latency. Clear the slot in result() and on every new scenario() so state can't leak across.
• Falls back to the old now - duration path when no start_time is supplied (BehaveX JSON replay path from fix: preserve behavex absolute timestamps when replaying JSON #494, third-party callers).
• Bumps qase-behave to 3.2.2.

Test plan

• Added TestParseStepRealTimeStart in test_utils.py (3 cases): explicit start anchors end = start + duration; explicit start excludes callback latency; fallback path preserved when no start_time.
• Added TestFormatterStepCallback in test_formatter.py (3 cases): step() stores wall-clock; result() forwards the stored value to parse_step and clears the slot; scenario() resets the slot to prevent leakage.
• pytest qase-behave/tests/ -v → 86 passed locally (Python 3.12).
• End-to-end verified against TestOps project DEVX with a 3-step scenario whose environment.py sleeps 200 ms in after_step:
  • Before fix (run #886): first step shows start_time +0.204 s relative to scenario start (200 ms drift); end_time is also +0.257 s (200 ms past real step end).
  • After fix (run #887): first step start_time is +0.0003 s relative to scenario start; end_time is +0.055 s — matches the real 50 ms step body, drift removed.
• CI green on all supported Python versions.