[TOPI] Add winograd for mali

[merrymercy]

Add winograd for 3x3 kernel.

Layers in VGG16

Workload	count	original spatial pack (GFLOPS)	winograd (effective GFLOPS)
(1, 224, 3, 64, 3, 1, 1)	1	12.1900	x
(1, 224, 64, 64, 3, 1, 1)	1	23.5718	23.853230
(1, 112, 64, 128, 3, 1, 1)	1	26.2391	30.027172
(1, 112, 128, 128, 3, 1, 1)	1	26.6143	35.892291
(1, 56, 128, 256, 3, 1, 1)	1	28.1192	41.253374
(1, 56, 256, 256, 3, 1, 1)	2	28.3494	45.154551
(1, 28, 256, 512, 3, 1, 1)	1	28.8210	47.407165
(1, 28, 512, 512, 3, 1, 1)	2	28.7867	49.158753
(1, 14, 512, 512, 3, 1, 1)	3	20.3852	34.234504

End-to-End

vgg16: 1.64 -> 0.98 (second / image)
resnet18: 0.20 -> 0.17 (second / image)

[tqchen]

@merrymercy do you think if it is possible to generalize to batch_size > 1?

[tqchen]

It is of interest to bring it to other backends as well, so I opened an issue here #899

[merrymercy]

@tqchen It depends on the GEMM implementation. Now It works well for batch size 1~32 with almost the same effective GFLOPS.