[Edu] Add missing ops import to GPU puzzle 18 starter scaffold

The `problems/p18/p18.py` starter scaffold imported only `DeviceRef,
Graph, TensorType` from `max.graph`, omitting `ops`. The puzzle's tips
and solution instruct learners to complete the graph definition with
`output = ops.custom(...)`, so following the documented workflow raises
`NameError: name 'ops' is not defined` before the graph is ever built.
Sibling puzzles 17 and 19 already import `ops`; add it to puzzle 18 so
the starter scaffold matches the documented steps. The import line sits
inside the doc-included anchor, so the rendered "Code to complete"
snippet is corrected as well.

Found while investigating a reported puzzle 18 compile failure. That
report is a different bug — a kernel elaboration error from `TileTensor`
mutability drift in `problems/p18/op/softmax.mojo`, addressed separately
by the skeleton resync work. This PR fixes the adjacent starter-scaffold
paper cut, which the `*.mojo`-only skeleton sync check does not cover.

MODULAR_ORIG_COMMIT_REV_ID: 4a717262be189a0ba6c7d70cddcc467b176b358a

Author: conorbronsdon

problems/p18/p18.py

@@ -10,7 +10,7 @@
 from max.driver import CPU, Accelerator, Buffer, Device
 from max.dtype import DType
 from max.engine import InferenceSession
-from max.graph import DeviceRef, Graph, TensorType
+from max.graph import DeviceRef, Graph, TensorType, ops  # noqa: F401
 from numpy.typing import NDArray
 from scipy.special import softmax as scipy_softmax