LatentMoE: K3 sigmoid routing and Switch aux/z-loss

Route with σ(W_r x), Top-k(s+b), then L1-normalize over the selected set.
Add Switch/GShard aux and router z-loss into train_k3 and train_sft.
Wiki parquet URLs honor HF_ENDPOINT for mirrored downloads.

kda/models/k3_config.py

@@ -53,6 +53,8 @@ class K3Config:
     moe_d_ff: int = 96
     situ_beta1: float = 4.0
     situ_beta2: float = 25.0
+    moe_aux_loss_coef: float = 1e-2   # Switch/GShard N Σ f_e P_e
+    moe_z_loss_coef: float = 1e-3     # mean (logsumexp logits)^2
 
     kda_backend: str = "reference"