Avoid switching between scalar and SIMD in Affine_vpts.

This might avoid a scenario where math works out differently
between skvx and scalar.

Change-Id: If5561fc451d27114a79f7d12d3bee381719c1db3
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/638222
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>