libdrmtap 0.5.9

Added: drmtap_crtc_refresh(), the exact refresh of the captured CRTC

drmtap_display.refresh_hz is the whole-hertz vrefresh of the mode, so on Linux 5.9
and later the cinema and broadcast rates lose their fraction: 59.94 reads 60, 23.976
reads 24, and timings that are not a whole number (1280x1024 at 60.02, a 3840x2160
mode at 29.98) read as if they were. A consumer that paces capture to the panel needs
the real rate.
drmtap_crtc_refresh(ctx, &num, &den) returns it as the reduced fraction num/den
hertz, computed from the current mode of the CRTC the way the kernel computes vrefresh
(pixel clock over the horizontal and vertical totals, with interlace, doublescan and
vscan) but not rounded. It is the rate the mode is programmed to: the kernel keeps the
pixel clock in kHz, so a 1080p "59.94" mode at 148352 kHz reads 148352/2475
(59.94020 Hz), a few parts per million away from the nominal 60000/1001.

No connector is probed: once the CRTC is known it is one DRM_IOCTL_MODE_GETCRTC, so
it can be called again to follow a mode change. On a context opened with crtc_id 0
the first call also picks the CRTC a grab would pick, and keeps that choice: the grab
and this call now share one helper for the auto-selection. -ENODATA when the CRTC
has no mode (disabled); a CRTC that is only blanked keeps its mode and answers.
DrmTap::crtc_refresh() in the safe wrapper returns Result<Option<(u64, u64)>>,
None for a CRTC with no mode.

Measured on i915: 60/1 for 1920x1080@60, 6750000/112463 (60.0197 Hz) for
1280x1024 and 131375/4382 (29.9806 Hz) for 3840x2160@30, matching the mode timings
in debugfs; -ENODATA on a disabled pipe; the auto-selected CRTC on a crtc_id 0
context. The mode-to-fraction step is pure and unit-tested on CEA-861 timings (60,
59.94, 50, 29.97, 23.976, 119.88, 1080i, doublescan, vscan) and on the widest values
the mode fields allow.

Fixed: the wrapper compared errno by its x86 number

DrmTap::plane_rotation() (0.5.8) recognised "no rotation property" as -95, which is
ENOTSUP on x86 and ARM but not on SPARC, MIPS or PA-RISC, so there a plane without
the property came back as an Err instead of Ok(None). The wrapper now depends on
libc and compares ENOTSUP and ENODATA by name.