Missing input data for a task should fail worker, not the task

If an input to a task is not actually in `self.data`, this currently manifests as the task failing (as though it's a user error). This in fact indicates a serious worker state issue, and should probably cause the entire worker to shut down.

See https://github.com/dask/distributed/issues/6125#issuecomment-1099877076 for a traceback of how an invalid transition led to a task executing where its inputs weren't all in memory yet. The `df.compute()` in client code failed as though this was a normal error, with the message `KeyError: "('split-shuffle-1-b4961b03aa9e8bec7c581d2dc337f717', 10, (3, 9))"`.

The problem is this overly-generous try/except in `Worker.execute`, which treats any problem in the `try` block as an issue with the task, including `_prepare_args_for_execution`, which looks up the input keys: https://github.com/dask/distributed/blob/68319f377da81057de8cfd4277f2448d796c592a/distributed/worker.py#L3488-L3601

Most problems that could happen there would be user error, but not all of them. We should be more selective?