fix: use raise instead of return for exceptions and avoid mutable default args

[Mr-Neutr0n]

Summary

• examples/pytorch/han/utils.py (line 266): Changed return NotImplementedError(...) to raise NotImplementedError(...). The original code silently returns the exception object instead of raising it, so unsupported datasets never trigger an error.

• examples/pytorch/evolveGCN/train.py (line 44): Changed return NotImplementedError(...) to raise NotImplementedError(...). Same silent-return bug — unsupported models are never flagged.

• examples/pytorch/bgnn/BGNN.py (line 212): Changed raise NotImplemented(...) to raise NotImplementedError(...). NotImplemented is a built-in constant (used for rich comparisons), not an exception class, so raise NotImplemented(...) throws a confusing TypeError instead of the intended NotImplementedError.

• examples/pytorch/jtnn/jtnn/jtnn_dec.py (line 83): Replaced mutable default argument clique=[] with clique=None and added a None guard. Mutable defaults are shared across all calls and can cause subtle bugs when mutated.

• examples/pytorch/jtnn/jtnn/chemutils.py (line 317): Replaced mutable default arguments prev_nodes=[] and prev_amap=[] with None and added None guards. Same mutable-default issue.

Test plan

• Verified each fix visually against the original source
• Existing tests should continue to pass — these are cosmetic/correctness fixes with no behavioral change for correct inputs