Make loona-hpack use RollMut/Piece

[fasterthanlime]

It uses &[u8], Cow<[u8]> and Vec<u8> respectively, when it should use Piece and others.

Ideally there'd be a way to make it generic over buffer strategies but... for now I'm not sure.

I'm not sure how that all plays with "provided buffers", or whether fluke will end up using them at all with io-uring, so it might be wise to keep this for a bit later.

Just... in benchmarks, I expect allocations to show up since fluke-hpack currently straight up allocates Vecs when it needs to do huffman decoding (see HuffmanDecoder::decode)