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.

data/eval/en2zh.src.txt

@@ -0,0 +1,20 @@
+The weather is very nice today.
+Please open the window.
+The cat is sitting on the mat.
+This book is worth reading.
+He went to Beijing yesterday.
+We need more training data.
+The sun rises in the east.
+She is studying linear algebra.
+Do not train on the evaluation set.
+There is an old bridge across the river.
+Science is the systematic exploration of nature.
+They are walking in the park.
+This computer has sixteen gigabytes of memory.
+Do not copy the source text when translating.
+The spring wind is gentle.
+I left the keys on the table.
+The train arrives at noon.
+Water boils at one hundred degrees Celsius.
+A small model can still learn a narrow task.
+Keep the frozen test files read-only.