Avoid marshaling unused bytes from the last uint64 in the bitVector to reduce bit overhead

[meling]

We could replace the []uint64 bit vector implementation we currently use with a []byte implementation. This can save up to 7 bytes per bit vector in unlucky cases, where the bit vector must use one additional uint64 when the vector's size isn't divisible by 64.

Note, however, that initial experiments with a []byte implementation seem to indicate that it is slightly slower to call bv.Set(). I have not tested it in BBHash.

We should benchmark the two implementations and compare them.

[meling]

We could consider letting the bit vector implementation be decided at compile time via generics. We would need a different implementation of words() to allow this.

[meling]

A better idea would be to strip off the parts of the last uint64 in the bit vector when marshaling, and correspondingly for unmarshaling.

That is, when marshaling we are converting from type bitVector []uint64 to []byte. In this process, when we reach the last uint64 in the bit vector slice, we should only marshal the necessary bytes from the last uint64.

This would avoid having to reimplement the bitVector as []byte to get the same savings.