`ref_sol` autograder pattern isn't inherently safe

A commonly-used pattern in autograders is the following:
```
def ref_sol():
###CORRECT_CODE###
return x

ref_x = ref_sol()

feedback.check_numpy_array_allclose("x", ref_x, x)
```
A trivial mistake (an extra `#` in `###CORRECT_CODE###`) caused the correct code to **not** be inserted in a situation that recently came up. As a result, `ref_sol` just returned the data from the surrounding scope, i.e. the solution provided by the user. In the example above, this leads to any result `x` being accepted, as long as it is a numpy array.

We should mitigate this problem on the Relate end. My favorite thought for now is that the `ref_sol` gymnastics could be automated, running a function containing the correct code in a clean scope, avoiding leakage and repeated manual implementation of the pattern.

x-ref: https://gitlab.tiker.net/teaching/cs450-f21/-/merge_requests/47#note_53515 (private access only, sorry)

Search the repository for ref_sol and ###CORRECT_CODE###, then trace how autograder reference solutions are assembled and evaluated. Reproduce the missing-insertion case and verify that reference code runs in a clean scope without falling back to the learner's surrounding variables; the feedback.check_numpy_array_allclose example defines the reported failure.

Written by the indexing model from the issue text.