This parameter is hidden and doesn't need to be supplied by users.
It allows to check index version and use different encodings/
decodings depending on the version.

This commit updates the vector encoding and decoding logic to use
`java.nio.ByteBuffer`. Using `ByteBuffer` shows an improvement in
[microbenchmarks](jtibshirani#3) and I
think it helps code readability. The performance gain might be due to the fact
`ByteBuffer` uses hotspot intrinsic candidates like `Unsafe#getIntUnaligned`
under the hood.

This commit updates the dense vector functions like `cosineSimilarity` to
decode the document vector and compute the result at the same time. Previously,
we would fully decode the vector into an array, then calculate the function.

This commit updates all dense vector functions to use `float[]` as opposed to a
`List<Number>` to track the query vector. The `float[]` query vector is held in
a new superclass `DenseVectorFunction`.

It also factors out the vector length validation into the superclasses
`DenseVectorFunction` and `SparseVectorFunction`.

This commit updates normalizes the query vector to unit length when
constructing `CosineSimilarity`. Since the query is already normalized, we
don't need to divide by its magnitude when computing the cosine.

…brute-force