#487 PCIe probe: median of several primed bursts instead of one

描述

`probe_pcie_h2d_gbps()` took a single 256 MiB x4 pinned burst after one warmup. On one x16 PCIe 4.0 link the same binary read 18.5, 6.9 and 5.8 GB/s across starts, and the low samples pin `pcie_frac` at 0.12-0.15 where calibration picks 0.35 (see #485).

This primes the link with two untimed copies and returns the **median of five** timed bursts, so one contended or not-yet-ramped sample cannot set the load's share. A >= 20 GB/s link still yields `pcie_frac 0.55`; a genuinely slow link now reports its true (low) rate instead of a worse outlier.

No new dependencies (`<algorithm>` is already included).

Refs #485.