Slight improvement (hopefully) to orientation sensing.

Since orientation sensing has been an issue for numerous users, I
decided a spend a little time experimenting with some possible
improvements.  I've settled on a couple major changes:

* Perform all lowpass filtering in spherical coordinates, not
  cartesian.  Since the rotations are what we're really concerned
  with, this makes more sense and gives more consistent results.

* Introduce a system of tracking "distrust" in the current data, based
  on external acceleration and on tilt.  The basic idea is after a
  signal of unreliable data -- repeated acceleration or
  nearly-horizontal tilt -- we wait for things to "stabilize" for some
  number of ticks before we start trusting the data again.  This is an
  extension of the basic lowpass filtering.  One simple example is
  after the phone is picked up off a table, we ignore the first few
  readings.  Another example is while the phone is under external
  acceleration for a while (i.e. in a car mount on a rough road), if a
  single "good" reading comes in, we distrust it, under the assumption
  that it was probably just a lucky reading (i.e. the magnitude
  happened to be close to that of gravity by chance).

These changes have allowed me to relax other constraints, such as the
filtering time constants, the maximum deviation from gravity, and the
max tilt before we start distrusting data.

The net effect is that orientation changes happen more quickly and can
happen under a wider variety of conditions, but false changes due to
tilt and acceleration are still avoided well.  I think the improvement
is subtle, but it's the best I've come up with in my limited time.

I've also included some refactoring and additonal comments to try and
further clarify the (somewhat twisted) logic.

Change-Id: I34c7297bd2061fae8317ffefd32a85c7538a3efb