graphics: harden weston-simple-egl against wall-clock steps

[Ricardo Salveti (ricardosalveti)]

weston-simple-egl fails intermittently on iq-9075-evk while every other test
on the same boot passes, including weston-simple-shm, KMSCube and core_auth,
with EGL running on real hardware (freedreno FD663).

Failing run:

Client finished, rc=143 elapsed=2885113s
FPS stats, samples=1471 avg=0.253841 min=0.200000 max=74.800003
[FAIL] Average FPS below detected-refresh threshold: avg=0.253841 < 64

Passing run, same test, same board:

Client finished, rc=143 elapsed=30s
FPS stats, samples=4 avg=74.850002 min=74.800003 max=75.000000

The difference is not the GPU. These boards have no valid RTC:

rtc-pm8xxx ...: setting system clock to 1970-01-01T00:00:16 UTC (16)

systemd bumps the clock to the image's build-time fallback, and NTP later
steps it 33 days to real time — that is the 2885113s. When the step lands
inside the 30s window the client's frame accounting jumps with it and prints
one single-frame report per frame; those flood the sample set and drag the
mean under the gate, even though the first window read 74.8 fps, the same
figure the passing run averages on a 75Hz link. Whether the correction lands
inside the window is timing, hence the flakiness.

Wait for the clock to settle before measuring, take elapsed from a monotonic
source, and check for a step afterwards. When one occurred, report SKIP
instead of asserting on numbers that cannot be right, and record the step
size and single-frame count in the summary.

Single-frame samples are counted, not filtered out: dropping them would hide
this flake but would equally hide a genuine 0.2 fps regression.

Tested:

• shellcheck -s sh -e SC1091,SC2230,SC3043 clean on both files
• FPS parser against captured logs: healthy log gives count=4 avg=74.9
  single=0; the 1471-line stepped log gives single=1470
• step arithmetic for no-step, the real 33-day forward step, a backward
  step, and malformed input
• tolerance validation rejects non-integer values

Runner/suites/Multimedia/Graphics/X11_GLX/run.sh uses the same FPS parser and
has the same exposure; left alone here for lack of failing data.

[Srikanth Muppandam (smuppand)]

LGTM