The 2/3 of validators would not censor by "not validating their txs," but rather by voting on/finalizing a chain that does not include the votes of these 1/3 of validators.
In this case, the amount that each set of validators get slashed is totally parameterizable. We could make it so the censoring validators lose exactly as much as the validators who are being censored. The relative amounts we punish here pretty much determine the greifing factors.
Note that censorship this is totally an issue in PoW, and to a much more concerning degree. In PoS, we can make it so censorship costs something; validators make the most money when all other validators are online. In PoW, if you have a >51% cartel, it's totally free to censor (and you can almost double your profits doing so). So note that while PoS has to worry about censorship, it's strictly better at handling it than PoW (as we could replicate PoW rewards in PoS).
If we're willing to accept large greifing possibilities, we can actually make it a dominant strategy for a cartel of any size to follow the protocol - which is something that is only possible in PoS.