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.

compose.yaml

@@ -0,0 +1,29 @@
+# GPU train / eval against the image in Dockerfile.
+#   docker compose run --rm train python train_k3.py --preset toy
+#   docker compose run --rm train swanlab ping
+#   docker compose run --rm train python -m kda.training.eval_mt --ckpt ...
+#
+# Secrets and corpora stay on the host.
+
+services:
+  train:
+    build: .
+    image: kda:latest
+    gpus: all
+    ipc: host
+    shm_size: "2gb"
+    working_dir: /workspace/kda
+    environment:
+      SWANLAB_API_KEY: ${SWANLAB_API_KEY:-}
+      HF_HOME: /cache/huggingface
+      HUGGINGFACE_HUB_CACHE: /cache/huggingface
+    volumes:
+      - ./ckpts:/workspace/kda/ckpts
+      - ./data:/workspace/kda/data
+      - ${PRETRAIN_DATA:-./data/pretrain}:/data/pretrain:ro
+      - ${EVAL_DATA:-./data/eval}:/data/eval:ro
+      - ${SFT_DATA:-./data/sft}:/data/sft:ro
+      - hf-cache:/cache/huggingface
+
+volumes:
+  hf-cache: