Fix #2946: Avoid evaluating T.inputs twice

[lolgab]

Fixes #2946

In evaluateNamed, after evaluating Tasks we are matching for ImputImpl[_] instances to construct a Watchable.
A Watchable instance contains a hashCode for the current result and a callback to evaluate it again and get the next result.
Instead of using the current result which was contained in TaskResult, we were already calling the recalc callback, which executes again the T.inputs and performs their side effects twice.
Now we are returning the hashCode for the already computed result, so we don't execute the recalc callback right away.

Pull Request: #2952

[lihaoyi]

Looks reasonable I think. @lolgab can you update the PR description with your understanding of the problem and your proposed solution before merging it?