Deprecate ImageQt align8to32() in favour of raw encoder stride

[akx]

Fell out of #9938 (namely 64507dc): the raw encoder had supported strides all along, but no one knew because it wasn't really documented in any way. For that reason, presumably, #1430 added a Python-side realigner.

This shows the old and new impls matching for a 113-pixel-wide Hopper, and what would happen if we don't align things:

[radarhere]

Hi. You're removing a public method. The preference would be to deprecate it instead.

[akx]

@radarhere Sure. Adjusted to a deprecation.

[radarhere]

For the record, here's a link - https://doc.qt.io/qt-6/qimage.html

each scanline of data in the image must also be 32-bit aligned

[radarhere]

I'm guessing this improves performance?

[akx]

I'm guessing this improves performance?

It does! +68% to +442% OPS for an image that wasn't correctly aligned on my box, ±10% (so measurement noise) for others. More than that, less memory allocations (as #9938 in general) since there isn't a temporary un-strided copy, temporary strided rows in a list, and finally a strided byte buffer, before the temporaries can be dropped.

[radarhere]

I've created akx#23 with some suggestions.

[akx]

Rebased to fix conflicts in the .rst files and folded in the changes from review + suggestion PR.