Possible imprecision in cast_shape

I'm not sure if this is a real issue or just the reality of floating point precision, but I wanted to at least flag it as suspicious.

I have a simple scene with a floor collider of shape
```rust
ColliderShape::cuboid(500.0, 1.0, 500.0)
```

I do a vertical sphere cast from above the floor via
```rust
fn test_sphere_cast(
q_pipeline: Res,
q_collider: QueryPipelineColliderComponentsQuery,
){
let collider_set = QueryPipelineColliderComponentsSet(&q_collider);
let shape = Ball::new(1.0);
let start = Isometry::translation(0.0, 3.0, 0.0);
let direction = vector![0.0, -1.0, 0.0];
let max_toi = 2.0;
let groups = InteractionGroups::all();
let filter = None;

if let Some((handle, hit)) = q_pipeline.cast_shape(
&collider_set, &start, &direction, &shape, max_toi, groups, filter
){
let y1 = hit.witness1.y;
let y2 = start.translation.y + hit.toi * direction.y + hit.witness2.y;
info!("{} {} {}", hit.toi, y1, y2);
}
}
```

I expect y1 and y2 will both agree on a hit height of ~1.0. In fact, y1 is always within about 1E-7. However, as I vary `start.x` and `start.z` in small 0.01 increments I notice that y2 may drift as little as 5E-8 or as much as 6E-2. I don't know the inner workings of the shape cast algorithm, but is this variation expected?