This PR adds a new module common::bitpacking with two structs: BitPacker and BitUnpacker. They would be used for the HNSW graph links compression in subsequent PRs.

Comparison with other implementations:

• bitpacking: it's fast, SIMD-optimized, but application is limited: blocks should be of size of 32/128/256 integers, all stored integers within a block should have the same bitsize.
• bitstream_io: the idea is somewhat similar to this PR, but this crate resulted to be slower for links compression/in-memory use-case.
• tantivy_bitpacker: optimized for a random access reads within a packed array.

The most important function is BitUnpacker::read() as it would be used in a hot loop. The inner buffer size is 8 bytes (a single u64) and it's optimized to read by whole 8 bytes at once.

The intended use-case for this module would look like the following (example numbers are arbitrary):

• Store header: 5 bit.
• Store first 32 values, each of size 12 bits. (e.g. sorted/delta-encoded for the main graph).
• Store last 20 values, each of size 15 bits. (e.g. additional links).
  Total size of the example: $\lceil \frac{5 + 32 \times 12 + 20 \times 15}{8} \rceil = 87 \ \text{bytes}$.