Optimize layer traversal

I found when investigating traces with reported jank from P there
was a large amount of time in handleMessageInvalidate but before
latchBuffer. The only culprit here could be layer traversal itself.
The current algorithm is quite costy requiring producing copies
of sp, allocating a new vector, etc (see makeTraversalList). This CL
optimizes the case where there are no relatives of the current layer.
I found that this does in fact eliminate the new time in
handleMessageInvalidate and also reduces onPreComposition.

Bug:69852584
Test: Manual, comparison of systrace, relative Z still works.
Change-Id: I2d942c1987087aae85decc0480d3622204ed665d