Release GNN statistics before the next message-passing hop (#1038)

> Release GNN statistics after aggregation so the previous hop’s tensor
does not overlap the next allocation. Model arithmetic is unchanged.
>
> Pretrained RelBench inference: 12 cases on an NVIDIA L4, PyTorch 2.14;
all predictions bitwise identical.
>
> | Largest classification case | Before | After |
> |---|---:|---:|
> | Transient prediction peak | 223.8 MB | 160.2 MB (−28.4%) |
> | Total allocated peak | 543.3 MB | 479.8 MB (−11.7%) |
>
> Latency was unchanged. A separate two-hop GPU Inductor check preserved
old/new outputs exactly; compiled transient allocation was unchanged at
7.07 MB because the compiler already eliminated the overlap.

Author: aw471

sdm/models/kumo/relational/invariant_gnn.py

@@ -109,6 +109,8 @@ def forward(
                 stats.flatten(1),
                 self.aggregation_lin.weight.T,
             )
+            # Release before allocating the next hop's statistics.
+            del stats
 
             if i == num_hops - 1:
                 start = graph.start_node_offsets[readout_table]