One major consideration is that the proof of work, as it's currently defined, has zero dependence on outside data.  This gives it a certain robustness, whereas if there were some dependence on a job list or some kind of computing grid external data, then the security of the entire system is degraded to the security of the external data source.  There is no guarantee that the jobs are as hard as they need to be, or that the system can't somehow be gamed or cheated.  The current double-SHA256 is impossible to cheat without breaking SHA256.

Better to just buy and sell compute cycles using bitcoins, rather than try to fit useful computations into the block chain proof-of-work.