flash: Optimize smem alloc for tcore for 8banks

Divide into first half & last half for warpgroup 0 & 1, and
allocate Q/K and P/V in different banks for parallel acccess.

tests/regression/flash_attention/flash_impl.hpp

@@ -11,8 +11,8 @@
 #define ROW_REMAINDER_LOGIC
 
 constexpr uint32_t ROWMAX_SETS = 3;
-constexpr bool WARP_SPECIALIZED = false;
-constexpr bool TENSOR_CORE = false;
+constexpr bool WARP_SPECIALIZED = true;
+constexpr bool TENSOR_CORE = true;
 
 // temporary safety stop for wrong configs
 static_assert(NUM_CORES == 4);