The .enable_vblank and .disable_vblank hooks were guarded by
EVDI_HAVE_CRTC_ATOMIC_STATE_ARG, which is false on 7.2 and later where the
callbacks moved to struct drm_atomic_commit. On those kernels the hooks
were dropped and vblank could never be enabled. Guard them with
EVDI_HAVE_IOSYS_MAP or EVDI_HAVE_DMA_BUF_MAP (5.11+) instead, the same
condition evdi_drm_drv.c uses for its driver level vblank hooks. The hooks
still return 1 here so this change is behaviour neutral on its own.

evdi faked vblank by returning 1 from enable_vblank and never delivering
timing to userspace, so compositors had no frame pacing. Drive
drm_crtc_handle_vblank() from an hrtimer whose period comes from the mode
timings (vtotal * htotal * 1e6 / clock, 60 Hz default). Start the timer in
enable_vblank, cancel it in disable_vblank and retune the period on mode
change.

Use div_u64 for the period so the 64-bit divide links on 32-bit ARM, and
READ_ONCE/WRITE_ONCE around vblank_period which is written on mode change
and read from the timer callback. Detect hrtimer_setup() via the new probe
and fall back to hrtimer_init() plus .function before 6.13. Cancel the
timer in evdi_drm_device_release_cb too and guard evdi_crtc_destroy
because drmm cleanup frees evdi before drm_mode_config_cleanup runs.

The painter used to hold the page-flip event until userspace grabbed
pixels, then sent it immediately, so flips completed off any real vblank.
Arm the event with drm_crtc_arm_vblank_event() so it fires on the next
tick, falling back to an immediate send if drm_crtc_vblank_get() fails.
Take event_lock with spin_lock_irq because the vblank hrtimer callback
takes it in hard irq context. Drop the now unused painter vblank path.

Cover the period math (default 60 Hz, a 144 Hz mode and a clock of 0
keeping the previous period), that the timer arms and is scheduled within
one period after drm_crtc_vblank_on and that it stops and the vblank count
freezes after drm_crtc_vblank_off.