vello_gpu: Replace image_bicubic with extended_images

[laurenz-canva]

As was shown in #1967, with my Android phone, I'm now able to get 90 FPS with most images! 🎉 However, that's not the case yet with my very old Android tablet. :( Currently, I'm still sitting at 22FPS, while Canvas2D manages to get a smooth 60FPS. This means there is more juice to be squeezed!

This PR proposes to improve that by replacing the image_bicubic feature with a extended_images feature. In addition to compiling out bicubic sampling, we also exclude any non-native image sampling path. This means if the feature is disabled, we also stop supporting:

• The upload_image method, as we can't support the internal image atlas.
• By extension, also glyph caching. This is a bit unfortunate, but I'm very certain that we can make this work with the native image sampling path in the future, but in order to avoid having too many changes at once, disabling it is easier for now.
• Any sample rect from an external texture that doesn't cover the whole texture.

Basically: Users can still render normal images with external textures using NN/bilinear sampling + any arbitrary extend mode (which I think covers the most common cases), but not more than that. As mentioned, the glyph caching situation can be improved in the future.

This gives a decent FPS boost on my tablet, though we are still far from reaching our target FPS.

With only #1968, I get around 22FPS:

IMG_1341.MOV

With this PR, I get around 31FPS:

IMG_1339.MOV

[LaurenzV]

It would be better to just compile out this method if the feature is disabled, instead of panicking at runtime.

However, internally we currently have a runtime feature flag, so unless we are willing to go "all-in" on external texture and remove that, we need to keep the method available, even if the feature is disabled.