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/_fla/ops/cp/__init__.py

@@ -0,0 +1,13 @@
+"""Context-parallel stubs. Pass ``cp_context=None`` (the default)."""
+
+
+class FLACPContext:
+    cu_seqlens = None
+    cu_seqlens_cpu = None
+
+
+def build_cp_context(*args, **kwargs):
+    raise RuntimeError("Context parallel is not included in the vendored KDA kernels")
+
+
+__all__ = ["FLACPContext", "build_cp_context"]

kda/_fla/ops/cp/chunk_delta_h.py

@@ -0,0 +1,11 @@
+"""Context-parallel hooks referenced by KDA fwd/bwd. Not implemented here."""
+
+
+def _cp_unsupported(*args, **kwargs):
+    raise RuntimeError("Context parallel is not included in the vendored KDA kernels")
+
+
+chunk_gated_delta_rule_fwd_h_pre_process = _cp_unsupported
+compress_h0 = _cp_unsupported
+chunk_gated_delta_rule_bwd_dhu_pre_process = _cp_unsupported
+expand_h0 = _cp_unsupported