refactor: namespace csrc kernels and extract common helpers

- attention family -> astrai::attention; fp8 family -> astrai::fp8
- new common/reduce.cuh (warp/group reductions, atomic_max_float)
- new common/cp_async.cuh (predicated cp_async_16, commit/wait group)
- move MAX_SPLITS into attention/common.h; delete warp_utils.cuh
- .cu bindings and pure C tests open family namespaces via using

csrc/kernels/attention/common.h

@@ -1,5 +1,10 @@
 #pragma once
 
+// Pure POD header
+
+namespace astrai {
+namespace attention {
+
 // Tensor layout for Q/K/V tensors passed to attention kernels.
 // Internally, kernels always operate on BHLD [batch, n_heads, seq_len, head_dim].
 // When the caller passes BLHD, dims 1 and 2 are transposed at entry.
@@ -8,6 +13,9 @@ enum TensorLayout : int {
     BLHD = 1,  // [batch, seq_len, n_heads, head_dim]
 };
 
+// Split-KV workspace cap: max decode splits per (batch, q_head).
+constexpr int MAX_SPLITS = 32;
+
 
 // Unified attention params covering BOTH addressing modes:
 //   - Contiguous K/V: dense [batch, kv_head, kv_len, head_dim] tensors (k/v).
@@ -73,5 +81,7 @@ struct AttentionParams {
     int num_splits;
     AT* __restrict__ o_part;
     AT* __restrict__ ml_part;
-
 };
+
+}  // namespace attention
+}  // namespace astrai