fix: Review fixes for end-to-end training pipeline

- Snapshot best-epoch weights during training and restore before
  checkpoint/RVF export (prevents exporting overfit final-epoch params)
- Add CsiToPoseTransformer::zeros() for fast zero-init when weights
  will be overwritten, avoiding wasteful Xavier init during gradient
  estimation (~2*param_count transformer constructions per batch)
- Deduplicate synthetic data generation in main.rs training mode

Co-Authored-By: claude-flow <ruv@ruv.net>

rust-port/wifi-densepose-rs/crates/wifi-densepose-sensing-server/src/graph_transformer.rs

@@ -452,6 +452,23 @@ impl CsiToPoseTransformer {
             config,
         }
     }
+    /// Construct with zero-initialized weights (faster than Xavier init).
+    /// Use with `unflatten_weights()` when you plan to overwrite all weights.
+    pub fn zeros(config: TransformerConfig) -> Self {
+        let d = config.d_model;
+        let bg = BodyGraph::new();
+        let kq = vec![vec![0.0f32; d]; config.n_keypoints];
+        Self {
+            csi_embed: Linear::zeros(config.n_subcarriers, d),
+            keypoint_queries: kq,
+            cross_attn: CrossAttention::new(d, config.n_heads), // small; kept for correct structure
+            gnn: GnnStack::new(d, d, config.n_gnn_layers, &bg),
+            xyz_head: Linear::zeros(d, 3),
+            conf_head: Linear::zeros(d, 1),
+            config,
+        }
+    }
+
     /// csi_features [n_antenna_pairs, n_subcarriers] -> PoseOutput with 17 keypoints.
     pub fn forward(&self, csi_features: &[Vec<f32>]) -> PoseOutput {
         let embedded: Vec<Vec<f32>> = csi_features.iter()