feat: Add Envar for Setting Pytorch Expandable Segments (#584)

**Reason for Change**:
Setting `PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True` allows
PyTorch Memory Allocator to mark memory blocks as expandable for reduced
peak GPU memory usage.

---------

Signed-off-by: Ishaan Sehgal <ishaanforthewin@gmail.com>

pkg/tuning/preset-tuning.go

@@ -348,6 +348,11 @@ func CreatePresetTuning(ctx context.Context, workspaceObj *kaitov1alpha1.Workspa
 			Value: "k_proj,q_proj,v_proj,o_proj,gate_proj,down_proj,up_proj",
 		})
 	}
+	// Add Expandable Memory Feature to reduce Peak GPU Mem Usage
+	envVars = append(envVars, corev1.EnvVar{
+		Name:  "PYTORCH_CUDA_ALLOC_CONF",
+		Value: "expandable_segments:True",
+	})
 	jobObj := resources.GenerateTuningJobManifest(ctx, workspaceObj, revisionNum, tuningImage, imagePullSecrets, *workspaceObj.Resource.Count, commands,
 		containerPorts, nil, nil, resourceReq, tolerations, initContainers, sidecarContainers, volumes, volumeMounts, envVars)