Use non-allocating versions of hex decoding

While working on initial tooling for #87, I found a surprising critical-path in our startup-from-cache path: we're spending over 13% of our CPU on hex decoding.

The likeliest factor here is tor-netdoc is using hex::decode, which allocates to the heap. We'd probably be much better off using hex::decode_to_slice.

It's an easy way to speed up a little bit, so I'll do it now.