Add bfloat16 support in RCCL

Preprocessor symbol RCCL_BFLOAT16 is used as feature indicator

src/enqueue.cc

@@ -30,7 +30,8 @@
   NCCL_FUNC4(coll, op, u64), \
   NCCL_FUNC4(coll, op, f16), \
   NCCL_FUNC4(coll, op, f32), \
-  NCCL_FUNC4(coll, op, f64)
+  NCCL_FUNC4(coll, op, f64), \
+  NCCL_FUNC4(coll, op, b16)
 #define NCCL_FUNCS3B(coll, op) \
   NCCL_FUNC4(coll, op,  i8), \
   NCCL_FUNC4(coll, op,  i8), \
@@ -40,6 +41,7 @@
   NCCL_FUNC4(coll, op,  i8), \
   NCCL_FUNC4(coll, op,  i8), \
   NCCL_FUNC4(coll, op,  i8), \
+  NCCL_FUNC4(coll, op,  i8), \
   NCCL_FUNC4(coll, op,  i8)
 
 // Must be consistent with ncclRedOp_t -- but we only generate kernel for sums.