Handle explicit alignment in `BufferVec`

`BufferVec` does not currently handle explicit alignment, and instead tightly packs data inside the buffer. This is fine for _e.g._ vertex buffers where there's no alignment requirement, but for other kinds of buffers this breaks the required hardware alignment and prevents the use of the buffer (see djeedai/bevy_hanabi#26 for an example of crash).

See _e.g._ #4319 for some example of alignments per hardware for uniform buffers. Storage buffers have similar requirements, highlighted by the [`WgpuLimits::min_storage_buffer_offset_alignment`](https://docs.rs/bevy/latest/bevy/render/render_resource/struct.WgpuLimits.html#structfield.min_storage_buffer_offset_alignment) which can be up to 256 bytes on macOS for example.

The proposal is to add an explicit `item_align` argument to the `new()` method, that the caller can fill with the expected alignment based on the buffer usage. Internally, there are multiple options for handling this:

1. Continue tightly packing the items into the internal `Vec`, and expand those into a properly-aligned CPU buffer at the point where the data is about to be copied to the GPU buffer. This has the advantage of lower CPU-side memory consumption, at the price of an extra memory copy.

2. Properly align the items with the `item_align` parameter when they're gathered in the CPU-side buffer (the `Vec`). This wastes the alignment padding, but avoids an extra allocation and memory copy when the GPU buffer is initialized.