Fix character controller ground detection attempt 3

[Vrixyz]

Following details from #711 (comment)

• always call detect_ground (which was not done if no collision were initially detected, or if we were not moving).
• flat value for predict_ground function

The tests were not passing though, a ground of 600 makes the controller not grounded at

  ┌             ┐
  │   502.99017 │
  │ -0.19247428 │
  │   100.54156 │
  └             ┘

By adding more height to the ground, I could make the test pass.

⚠️ This PR poses a problem of detecting ground when too near of a ground, because loosened shape intersect with it.

video illustration

Unfortunately we can´t see in the video, but during a frame the charactercontroller is detected as grounded when flying exactly on the edge of the "cliff":

Screencast.from.09-09-2024.04.21.37.PM.webm

It's not the case before because we were not computing the ground state when there was no collision on the initial cast_shape.

[sebcrozet]

Looking good, thanks.
It’s interesting that you sometimes get minY ~= 0.02. I’m wondering if the offset gets counted twice somewhere (because 0.02 ~= 0.009 * 2 + 0.004).