The dw_hdmi enable/disable handling is particularly weak in several
regards:
* The hotplug interrupt could call hdmi_poweron() or hdmi_poweroff()
  while DRM is setting a mode, which could race with a mode being set.
* Hotplug will always re-enable the phy whenever it detects an active
  hotplug signal, even if DRM has disabled the output.

Resolve all of these by introducing a mutex to prevent races, and a
state-tracking bool so we know whether DRM wishes the output to be
enabled.  We choose to use our own mutex rather than ->struct_mutex
so that we can still process interrupts in a timely fashion.
Signed-off-by: NRussell King <rmk+kernel@arm.linux.org.uk>