read_restart: use `time_bnds` instead of `time_bounds` for CF compatibility

[Chris Black]

This seems like a reasonable fix for this PR. Long term it might be worth being consistent about calling this time_bnds throughout, which I think is better aligned with the CF standard: ORNL DAAC lists time_bnds as required, and time_bnds appears* in the conventions themseves whereas time_bounds does not. Can we consider a followup patch to make read_restart use time_bnds?

*Though admittedly time_bnds only appears in an non-normative example. The CF conventions are messy!

Originally posted by @infotroph in #3249 (comment)