In-memory representation of slice run iterator is inefficient

[Stebalien]

The current in-memory representation of the run iterator is really inefficient. On a 64 bit system, it'll take 16 bytes per run due to alignment. This is an issue because we cache this iterator whenever we first try to read a bitfield.

We can cut this in half by:

1. First combining (or failing) on adjacent runs with the same value.
2. Encoding the runs as a slice of 64 bit integers where the value if 0 elements are even, and the value of odd elements are 1.

We can make this even smaller if we're willing to use a variable-length integer encoding (or simply avoid caching and get better at decoding in-place).

[acruikshank]

sorry, closed it accidentally.