perf(pixmap): unify the registry on get() and bound the caches by bytes

[gmmcosta15]

Description

• Feature
• Bug fix
• Code refactor
• Documentation

Last PR of the BlocksPixmap stack. Collapses the registry to a single entry point and
replaces the entry-count caps with byte budgets.

• scaled() folded into get(): one call takes an Icon or an existing QPixmap, and any QSize/QSizeF/QRect/QRectF as the target.
• _as_size() normalises the target in one place; 4 call sites move from truncate to round (1px max).
• The three caches are byte-bounded LRU (_pixmaps 12 MiB, _scaled 8 MiB, _tints 8 MiB) instead of entry-capped.
• tinted() takes a composition mode, so blocks_button stops allocating a pixmap per paint.
• source() added for the three call sites that need the intrinsic size.
• Icon.LOGO_BLOCKS replaces a raw filesystem path that resolved to nothing on device.
• Drops Antialiasing + SmoothPixmapTransform from blocks_button._paint_pixmap.

Motivation

An entry cap cannot bound a mixed-size cache. A 64px icon rescale is 16 KiB and a 400x300 thumbnail is 480 KiB, so a 64-entry cap ranges 2-60 MiB purely on how the user browses. Qt bounds QPixmapCache by bytes for the same reason. On a 2 GB box with 209 MiB free that variance is the whole budget.

The two render hints are no-ops on a 1:1 integer-offset blit, which is the only thing _paint_pixmap does now that the scale happens in the cache.

Tests

New coverage: byte-bounded eviction on both caches, LRU order (a FIFO would drop the startup icons first), aspect-mode and colour key separation, and every QSize/QSizeF/QRect/QRectF target normalising to one cache entry.