Improve allocations in encoders

[RunDevelopment]

Small PR. Changes:

1. BCn encoder don't allocate for each line anymore. If the width of an image encode with a BCn format isn't divisible by 4, we have to add padding to get a full 4x4 block. This full block was previously allocated on the heap for every row of the image. I fixed this by adding a new const generic param BLOCK_SIZE (= BLOCK_WIDTH * BLOCK_HEIGHT) to use an array on the stack instead.
2. For bi-planar formats, the vec for the 2nd plane now allocates all its memory up front instead of growing dynamically.
3. Where possible, I converted vecs into boxed slices to make accidental reallocations impossible.

This makes generally makes it easier to reason about allocations and should make the code a tiny bit faster.