Embed ST actions in Eval

[treeowl]

I'm not sure just how useful this is, but it seems like it should exist.

[treeowl]

One question to consider: are there situations where the lazy blackholing here will probably lead to repeated computation? If so, we might just want to derive MonadFix from the IO instance and take the minor performance hit.

[konsumlamm]

Without any use case, I'm not convinced that this is a useful addition. Also, can't this always be implemented as stToEval m = runST (pure <$> m)? In that case, I'm even less convinced to add this, since it can be easily defined in user code.