Initial K3 snapshot: 0.5B KDA/MLA/MoE train path

Standalone tree split from LLMRL/projects/kda. Includes Triton dt_bias
backward fix, train_k3 --preset 0.5b, SFT, Docker runtime, and tests.

kda/layers/__init__.py

@@ -0,0 +1,23 @@
+"""Composable mixing layers: attn and ffn both map [B,T,D] -> [B,T,D].
+
+Depth mixing (AttnRes) is not a layer_specs kind. CausalLM reads
+``config.attnres`` (off | full | block) and wraps DecoderBlock sublayers.
+"""
+
+from .block import DecoderBlock, build_attn, build_ffn
+from .kda_attn import KDAAttention
+from .latent_moe import LatentMoE
+from .mla import GatedMLA
+from .rmsnorm import RMSNorm
+from .swiglu import SwiGLUMLP
+
+__all__ = [
+    "DecoderBlock",
+    "GatedMLA",
+    "KDAAttention",
+    "LatentMoE",
+    "RMSNorm",
+    "SwiGLUMLP",
+    "build_attn",
+    "build_ffn",
+]