cudaStreamEndCapture / cudaGraphInstantiate / cudaGraphDestroy were
guarded by assert(), which abort()s the whole process on failure. These
calls allocate GPU memory and fail with OOM when a trial's config sits
near the VRAM ceiling. Because graph capture happens during
create_pufferl_impl's warmup, a single memory-hungry sweep trial would
take down the entire sweep process rather than being skipped.

Throw std::runtime_error instead. The exception propagates up through
create_pufferl to the existing RuntimeError handler in pufferl.py, which
already skips a trial and continues the sweep (same path as the
"CUDA OOM: failed to allocate training buffers" case). Applied to both
the rollout (net_callback_wrapper) and train (train_impl) graph captures;
the graph is destroyed before throwing on an instantiate failure to avoid
leaking it.