Do all color-mapping on the GPU

[emilk]

Currently we color-map tensors and depth-images on the CPU. This has many downsides:

• It is slower
• It requires complex caching on the CPU
• It blocks Use less RAM: remove egui images from ImageStore #506

Instead, let's improve the re_renderer::TexturedRect API to handle:

• Color map selector enum
• A range for the color map
• Gamma
• Color maps for all 2D tensors in the image view
• Color maps for all tensors in the tensor view

Misc TODO

• Remove the extra enum Colormap that we have
• Support texture filtering

[emilk]

We have a few different cases to cover.

The first is no color mapping, e.g. a normal RGBA / texture_2d<f32> / Rgba8UnormSrgb texture.

Then we have single-channel color mapping of:

• texture_2d<u32> (R8Uint, R16Uint, R32Uint)
• texture_2d<i32> (R8Sint, R16Sint, R32Sint)
• texture_2d<f32> (R8Unorm, R8Snorm, R16Float R32Float)

Note that we have a choice for 8-bit integers if we want to normalize them (using texture_2d<f32>) or not (using texture_2d<i32> or texture_2d<u32>). For 16-bit and and 32-bit integers we don't have that choice, so we still need all three texture_2d types.

The non-color-mapped path should probably be separate, but for the color-mapped path we have the choice to either write three shaders, or combine them into one. For the latter we would use a runtime-switch to know what texture_2d to read from. The two unread ones would be bound to dummy textures.

[emilk]

Also: all images in our UI should be shown using the same re_renderer shader, and we should get rid of get_colormapped_view/ColoredTensorView