feat(cuda): add int4 ptx 256x64 mma kernel

GitOrigin-RevId: 8f7475b0f6f60c13f7cc8e571c81822d72b6f689

dnn/src/cuda/ptx/uint4_int4/imma8832_256x64.cuh

+#pragma once
+
+#include "./base.cuh"
+
+#define TX             128
+#define TY             1
+#define BM             64
+#define BN             256
+#define BK             128
+#define mma_m          16
+#define mma_n          8
+#define mma_k          64
+#define reg_m          8
+#define reg_n          8
+#define packed_channel 64
+#define BKd32          (BK / 32)
+#define reg_md4        (reg_m >> 2)
+#define WARPS          (TX / 32)
+#define cache_per_warp 128
+#define reg_nd4        (reg_n >> 2)
+#define ldg_src        (BN * BK / (16 * TX))
+#define ldg_filter     (BM * BK / (16 * TX))
+#define ldg_width      16
+
+// vim: syntax=cpp.doxygen