Faster par iter bp

[RagnarGrootKoerkamp]

Fix #15:

• Instead of re-initialising a Box<[u32x8; 8]> on each call, we use a thread-local buffer.
• Instead of read_slice_32 that has a bounds-check, we propagate the padding of the PackedSeqVec via as_slice, so that we can just always do unchecked reads. Big speedups from this for short seqs.
• Tried gather instead of transpose, but it's only marginally better. (See gather_instead_of_transpose branch.)