fix the handling of PartialResultSet.resume_token

[devbww]

We currently assume that the service only adds a resume_token when the PartialResultSet ends with a completed row (except when the PartialResultSet also completes the stream, in which case the resume_token remains empty).

While this appears to be true in practice (and we create a "PartialResultSetSource reader produced a resume token that is not on a row boundary" error when it is not), apparently we should allow for a resume_token in an incomplete row, and even in the middle of a chunked_value.

See #9359 for more context on the current state. Internally, see go/cloud-spanner-handling-resume-token.

[coryan]

Changed to P3 as it seems the problem is theoretical. It would be nice to fix, but nothing is broken at the moment.

[scotthart]

Will address this when we encounter issues.