Use self-descriptive cache with type tags

[ilevkivskyi]

This should make our cache format more future proof w.r.t lazy deserialization and other features:

• This makes cache ~5% larger (data + meta)
• This makes interpreted performance ~5% worse (just the serialization steps, not overall time)
• No visible performance impact when compiled (probably because all extra indirections like read_str() -> read_str_bare() etc are C calls).

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅