Improve text rendering

[akx]

Note

Stacked on #9908 (I guess GitHub's merge stacks don't work cross-fork, ah well).
That PR ought to be reviewed and merged first and I'll happily rebase this.
24a6fa0 is the first new commit here.

Refs #9898 (not "fixes", since I can't be sure this ends up fixing OP's issues).

What's moving?

Some infrastructure:

• PILLOW_TEST_SNAPSHOTS can now be set to create, overwrite or update, so assert_image_similar_tofile and assert_image_equal_tofile can update the golden image to what was passed in; this is similar to --snapshot-update for Syrupy or INSTA_UPDATE for Rust's Insta.
• A test based on Kerning woes #9898's example code, golden snapshots updated as they improve over the patches.

These are the actual fixes; see #9908 for notes on the clean-ups:

• Passing in a bytestring where some characters had the high bit set would render mojibake/tofu/.NOTDEF (whatever you want to call it) when using the basic renderer. Very corner-case, but also the fix is just a proper (unsigned char) cast.
• Fractional variable font axis values were truncated; no more!
• When using basic font layout (which is the case for the default font added in Added TrueType default font to allow for different sizes #7354), advance values were 64× off due to type confusion (see Document and clarify _imagingft #9908's description). No more!
• Finally (the last 2 commits), glyph positioning within a single call is now subpixel and not pixel-snap-hinted (when available), leading to more accurate rendering.