We prefer to prevent using chained assignment (`T a = b = c = T();`) as this
can lead to confusing code and subtle bugs.
According to https://en.wikipedia.org/wiki/Assignment_operator_(C%2B%2B), C++
allows any arbitrary return type, so this is standard compliant.
This could be re-assessed if/when we have an actual need for a behavior more
akin to that of the C++ STL, for now this PR simply changes a handful of
cases which were inconsistent with the rest of the codebase (`void` return
type was already the most common case prior to this commit).
Bounce calculation now uses the previous frame's velocity, so it's
consistent with the actual motion of the bodies involved and not the
yet-to-be-applied forces.
When bounce is 1, using the current velocity was causing the new forces
(including gravity) to be taken into account, which lead to the bounce
velocity to be higher than the falling velocity at the moment of impact,
adding more and more energy over time.
Regression fix, gravity was accumulated between frames after some
changes around area gravity calculation.
Also got rid of unused member and method in soft body class.
Replaced the previous implementation for backface collision handling (in
test_axis function from SAT algorithm) with much simpler logic in the
collision generation phase with face shapes, in order to get rid of
wrong contacts when backface collision is disabled.
Now it just ignores the generated collision if the contact normal is
against the face normal, with a threshold to keep edge contacts.
Added a special case for soft bodies to invert the collision instead of
ignoring it, because for now it's the best solution to avoid soft bodies
to go through concave shapes (they use small spheres). This might be
replaced with a better algorithm for soft bodies later.
Updating the broadphase to find new collision pairs was done after
checking for collision islands, so it was working in most cases due to
the pairing margin used in the BVH, but in case of teleported objects
the narrowphase collision could be skipped.
Now it's done before checking for collision islands, so we can ensure
that broadphase pairing has been done at the same time as objects are
marked as moved so their collision can be checked properly.
This issue didn't happen in the Octree/HashGrid because they do nothing
on update and trigger pairs directly when objects move instead.
