Combine RGBA8 linear/nearest sampling texture cache

[kvark]

Instead of having different texture arrays for nearest/linear sampling, we should just switch the sampler and use the same texture array. It will still break batches, but we could possibly use the texture space more efficiently.

In the future, we can go even further and pack the single-channel (and dual-channel, for the matter) data into the same RGBA8 texture array, as long as we provide a swizzle matrix for the shader to apply after fetching.
Note: there are downsides of this approach, and it is to be considered separately. Brought up here for the reference only.