Optimize XOR floating point compression

[cswinter]

Now achieves up to 390MiB/s encode, 1050MiB/s decode on random [0, 1] double precision values. Encode up to 415MiB/s for doubles created as [0, 1] single precision values.

Main changes:

• sub out simple BitWriter/BitReader implementations for https://github.com/icewind1991/bitbuffer, this is where majority of speed gain comes from.
• perform a single 2 bit write for the 10 and 11 control codes rather than two 1 bit writes (~40% speedup)
• when decoding, pre-allocate entire Vec and replace push with direct write. (~15% speedup)