Direct3D buffer-to-texture copies don't need a temporary buffer

[jimblandy]

To fix #4150, the direction we're pursuing in #7721 is to copy the unaligned data into a temporary buffer, properly aligned. But surprisingly, there is another way to fix this that requires no temporary and a constant number of copies.

Dawn implements this, with lots of comments. The 3D version is very challenging, but it can be done.

[Kangz]

Note that it will also need this piece of logic to work around D3D12 validation of the buffer size not matching WebGPU's exactly. The code is being updated to support BC sliced 3D in Dawn so it's probably best to check the top of tree when looking at implementing it.