Store _prev as a tuple in sum() and mean()

[eeshsaxena]

Fixes #7.

The bug

Every op builds its _prev through _create_child, which normalises the argument with tuple(prev). sum() and mean() skip that helper and assign the set directly:

out._prev = {self}

so those two nodes carry a set while every other node carries a tuple.

_eval_forward pulls operands by index:

elif node._op == "sum":
    values[node] = values[node._prev[0]].sum()

Indexing a set raises TypeError: 'set' object is not subscriptable, so evaluating any graph that reduces with sum or mean (which is almost every loss) crashes.

The fix

Store (self,) in both methods so _prev is a tuple everywhere:

out._prev = (self,)

Test

Added test_eval_forward_handles_reductions, which runs a sum graph and a mean graph through _eval_forward and checks the values (30.0 and 14/3). It raises TypeError on main and passes with this change. The full tests/test_gradcheck.py suite passes locally.

One heads-up while I was in there: grad_check has a separate, unrelated problem where it perturbs the reduced output node (node.data[idx] = ...) whose data is a scalar, which raises 'numpy.float64' object does not support item assignment. That is outside the scope of this fix, so I left it alone, but happy to open a follow-up if useful.

[Terminay]

Hi! Welcome to Leanpass, @eeshsaxena; this is a genuinely huge milestone for me. You are leanpass's FIRST contributor!!

[Terminay]

Good catch about grad_check turns out #17 doesn't actually cover the scalar item-assignment crash you found, so it's a separate thing after all. Could you open a new issue for it? Your write-up above is basically issue-ready as-is; feel free to just copy it over. Happy to have you pick it up too if you want, but no pressure either way!