Add the Kimi K3 backend primitives

[yueming-yuan]

Three hooks the Kimi K3 plugin in miles (radixark/miles#1825) relies on, ported from the K3 review fork (yueming-yuan/k3-megatron-review@ebe3b9d7f) onto miles-main. They were never opened against this repo; the K3 image carried the fork instead.

• TransformerConfig.gated_activation_func: an activation that consumes the packed [gate | linear] GLU input itself (K3's situ). MLP and TEGroupedMLP call it in place of activation_func(gate) * linear when set; None keeps today's behavior.
• TransformerConfig.moe_latent_use_norm: RMSNorm the combined routed output in latent space before fc2_latent_proj (K3's routed_expert_norm, a real checkpoint tensor).
• Parameter.sum_gradients_across_tp_domain: sum, not average, TP-replicated partial gradients in finalize_model_grads, for K3's expert-shared LoRA factors whose gradient is a partial sum over the TP shards.

Without the norm hook a K3 checkpoint loads with the latent norm silently dropped; --check-weight-update-equal on the 4-layer K3 run caught it as routed_expert_norm.weight never reaching the engine.

Validation

• 4-layer Kimi K3 LoRA RL smoke run (colocated, 8xH200, Kimi-K3 full-model and LoRA RL support miles#1825 + sgl-project/sglang yueming/k3-lora-on-sglang-miles) with --check-weight-update-equal: base weights compare equal after the sync, two rollouts train.
• Not executed: a TP>1 run exercising sum_gradients_across_tp_domain (the smoke run is TP=1); no unit test — the fork's only test covered an unrelated SGD change that is already on miles-main.