All in all, I think yes. The primary reason is that the silent switch to float introduces numerical instability in several places that is often hard to track down. In addition, we use double in a lot of places on non-embedded systems where float would be sufficient.
Some considerations though:
A lot of structures will now suddenly grow, QRectF, QPointF along with user-structs that use qreal. I don't think this will be a big issue, but it is worth mentioning.
Inside the QPainter API, everything will now turn to double precision, but this is mostly in the calling API, so probably not a big deal. The raster engine is pretty much all fixed point math internally. The GL 2 paint engine uses qreals, but the real work is done by the GPU, so also not a big deal.
Inside scene graph, we explicitly use floats already for our geometries, so no worries there.
The matrix and vector classes would require a bit of thought. QMatrix4x4 is qreal and precision is potentially an issue. Inside scenegraph and I believe also inside Qt3D, we have not had real issues with using only floats though, so it might be overkill for our usecases. These classes are for 2D and 3D graphics, not for generic math. There are better libraries for that available. QQuaternion is qreal for some reason while QVectorXD stores float internally even though all the API takes qreal. Maybe the entire API should be float based? Another option is to duplicate the matrix class using a template baseclass. One that uses float for scenegraph and one that uses doubles for high precision math.