fix: stabilize peak memory checks in HPU performance test
Peak RAM/VRAM in the perf pipeline flipped between PASS and FAIL with no
algorithm change (identical loss, tuning time and output size):
- Peak RAM: measured by a harness-owned probe (measure_peak.py) that runs
the auto-round CLI in-process and records the kernel RSS high-water mark
(ru_maxrss), so it no longer depends on per-block sampling timing or
short-lived child processes, and current/baseline share the same ruler.
- Peak VRAM: MemoryMonitor.update_hpu now synchronizes before sampling, so
tensors whose frees are still queued in the HPU eager pipeline are not
randomly counted. The allocator high-water mark is logged for info only
because on HPU it varies with host/device timing.
- Gate: |diff| <= max(relative, absolute) with absolute floors of 0.15GB
(RAM) and 0.25GB (VRAM); still two-sided.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Sun, Xuehao <xuehao.sun@intel.com>