Revert "Increase precision of some quad AA calculations"

This reverts commit 40a135312688beec0eb3aa74ac6612393c1c0a6d.

Reason for revert: layout tests need rebaseline

Original change's description:
> Increase precision of some quad AA calculations
> 
> This stops using the low precision rsqrt and instead does a divide + sqrt.
> It also separates the single kTolerance value used in GrQuadUtils' math
> into a fairly fuzzy tolerance for edges/distances being close to zero
> (e.g. a degenerate quad), and a more precise tolerance for linear systems
> solutions.
> 
> The smaller tolerance value used for denominators in these solutions prevents
> the solved-for coordinates from being mis-classified as "bad". The bad coord
> resolution works on the assumption that its correcting points that are
> co-located, but if the input coords were sufficiently large the linear
> equations would have small denominators that were less than the fuzzy tolerance
> and then the geometry would be improperly collapsed to a line.
> 
> For the selected zoom level of GM_crbug_10141204, this change is sufficient
> to prevent the draw from being discarded. However, when animating the zoom
> level, the input shape is sometimes still considered a fullscreen clear when
> it shouldn't be. A follow up change will improve precision in CropToRect()
> 
> Bug: skia:9779, chromium:10141204
> Change-Id: Id1518a908c03e7813868c4eff014d18388f8f62e
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/265976
> Reviewed-by: Robert Phillips <robertphillips@google.com>
> Commit-Queue: Michael Ludwig <michaelludwig@google.com>

TBR=robertphillips@google.com,michaelludwig@google.com

Change-Id: I44012a2728f75e235678f733c6da8870b84ae3ab
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: skia:9779, chromium:10141204
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/266220
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
1 file changed