Previously, we would decode the PEM wrapper for keys twice: once in
get_next_token, and once later in PEM decode.  Now we just do all of
the wrapper and base64 stuff in get_next_token, and store the
base64-decoded part in the token object for keys and non-keys alike.

This change should speed up parsing slightly by letting us skip a
bunch of stuff in crypto_pk_read_*from_string(), including the tag
detection parts of pem_decode(), and an extra key allocation and
deallocation pair.

Retaining the base64-decoded part in the token object will allow us
to speed up our microdesc parsing, since it is the asn1 portion that
we actually want to retain.