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/training/prompts.py

@@ -0,0 +1,7 @@
+"""Instruction strings shared by SFT and eval. Do not drift."""
+
+
+def instruction_prompt(src: str, target_lang: str) -> str:
+    if target_lang.startswith("zh"):
+        return f"Translate to Chinese:\n{src}"
+    return f"Translate to English:\n{src}"