Qt5Agg HiDPI support

[QuLogic]

This completes the work by @piannucci in #6389. Scaling is necessary on the figure DPI, and inverse scaling on rubberband and blitting locations.

I don't have a HiDPI display, but I tested by setting QT_SCALE_FACTOR=2. There is one bug with resizing if I use the grip in the status bar because it produces the wrong size. I'm not sure where that is created, so I haven't fixed it yet. Resizing using the OS's size works fine.

Ping @tacaswell, @astrofrog, @adrn, and @jenshnielsen who appear to have HiDPI screens based on other issues.

[astrofrog]

Seems to work nicely!

[tacaswell]

There is a bug with fig.set_size_inches, it uses the scaled DPI to resize the canvas.

[QuLogic]

I'm not seeing the WebAgg backend (#5383) do anything to handle that either; is it really needed?

[tacaswell]

The set_size_inches bug is probably related to the mouse-resize issue?

[QuLogic]

On resize events, we scale the logical pixels to physical pixels and then calculate inches via the scaled DPI, so I'm pretty sure fig.set_size_inches should use the scaled DPI.

Looking at the resizeEvent in FigureCanvasQT, I get logical sizes of ~480x80 when resizing using the OS handles. When using the grip in the toolbar, I get logical sizes of ~729x187, which aren't even double! I assume these values come directly from Qt as I see nothing else in the traceback.

So I'm wondering whether this bug occurs with true HiDPI displays? Is it a bug in Qt or just Qt's QT_SCALE_FACTOR?

[QuLogic]

Found QTBUG-51899 and QTBUG-53389, so this is not a bug in the matplotlib code.

[NelleV]

This looks good. I am not very familiar with these backends, so another pair of eye should look at this PR.

[MAKOMO]

Works for me as expected. Would be cool to have this in the final v2.0 already.

[jenshnielsen]

Tested it too and seems to work fine. The only question is if it's to late to merge into 2.0 @tacaswell