The current code in `SipHasher128::short_write` is inefficient. It uses
`u8to64_le` (which is complex and slow) to extract just the right number of
bytes of the input into a u64 and pad the result with zeroes. It then
left-shifts that value in order to bitwise-OR it with `self.tail`.

For example, imagine we have a u32 input 0xIIHH_GGFF and only need three bytes
to fill up `self.tail`. The current code uses `u8to64_le` to construct
0x0000_0000_00HH_GGFF, which is just 0xIIHH_GGFF with the 0xII removed and
zero-extended to a u64. The code then left-shifts that value by five bytes --
discarding the 0x00 byte that replaced the 0xII byte! -- to give
0xHHGG_FF00_0000_0000. It then then ORs that value with self.tail.

There's a much simpler way to do it: zero-extend to u64 first, then left shift.
E.g. 0xIIHH_GGFF is zero-extended to 0x0000_0000_IIHH_GGFF, and then
left-shifted to 0xHHGG_FF00_0000_0000. We don't have to take time to exclude
the unneeded 0xII byte, because it just gets shifted out anyway! It also avoids
multiple occurrences of `unsafe`.

There's a similar story with the setting of `self.tail` at the method's end.
The current code uses `u8to64_le` to extract the remaining part of the input,
but the same effect can be achieved more quickly with a right shift on the
zero-extended input.

All that works on little-endian. It doesn't work for big-endian, but we
can just do a `to_le` before calling `short_write` and then it works.

This commit changes `SipHasher128` to use the simpler shift-based approach. The
code is also smaller, which means that `short_write` is now inlined where
previously it wasn't, which makes things faster again. This gives big
speed-ups for all incremental builds, especially "baseline" incremental
builds.

This makes it faster and also changes it to a safe function. (Thanks to
Michael Woerister for the suggestion.) `load_int_le!` is also no longer
necessary.