The code in `siphash24` was obtaining the pointer in buffer at offset by
accessing an element at that offset instead of accessing a range or
simply computing the offset of the pointer from the start. This is UB
because one canot access past `T` even if the allocation is known to be
large enough. This change fixes it by using a range and also replaces
complicated code with simpler use of `from_le_bytes`.
It's quite likely that this can be improved further, possibly even
removing the `unsafe` without speed penalty but it's a larger task
that's not a priority right now.