The issue was caused by incorrect check for exceeded number
of transparent bounces: the same maximum distance was used
for picking up N closest intersections and counting overall
intersections count.

Now made it so intersection count is using ray distance which
matches the way how Embree and OptiX implementation works.

Additionally, T95462 is probably fixed with this change,
although some closer look is needed.

This is likely to cause some performance loss, but this is
unlikely to be avoided.

One thing I've notices is that the ray distance of FLT_MAX
will be "shortened" after instance push. Technically, this
can lead to missing intersection checks, but probably unlikely
to happen in practice. Can address it in a separate patch.