Issue with flattening of a stroked cubic around a near-cusp

[armansito]

On Apple hardware, the following near-cusp test case from the tricky_strokes test scene has this artifact at certain scale factors:

This specifically only seems to happen with the near-cusp defined as [(0., 0.01), (128., 127.999), (128., 0.01), (0., 127.99)]. The bug itself is in the flatten stage, specifically the cubic-to-es subdivision loop in flatten_euler. I debugged this over a couple days and here are my findings so far:

• For that specific cubic, the issue is only present for a (uniform) scale factor within the range 5.3332 and 113.5797. Scale factors outside that range do not exhibit the issue. This can be verified by hard-coding a scale factor right before testing the scaled error estimate against the error tolerance:

...
scale = 5.3332;
if cubic_params.err * scale <= tol || dt <= SUBDIV_LIMIT {
...

• The issue is not present when fastmath is disabled. The Metal compiler enables optimizations on floating-point arithmetic (which may violate the IEEE-754 standard) by default. I verified this by setting fastMathEnabled = NO on MTLCompilerOptions directly on Skia-Graphite's Metal backend (which uses the shaders directly and doesn't use wgpu).
• The issue is not present on CPU shaders (which doesn't use fastmath).