1.optimize add by channel; 2. texture reuse should consider param input and multi fetch; 3. should also realloc heap when max size become smaller; 4. paddle mobile test disable cache (#1769)

* optimize add by channel

* 1. texture reuse should consider param input and multi fetch; 2.  should also realloc heap when max size become smaller; 3. paddle mobile test disable cache

* add slice cpu op

* print error var name && print op position

* add exp cpu

* add exp cpu

* add comment

* add comment

* add comment

* add CPU_ARCH info, improve the performance of GEMM1*1s1

* improve the performance of gemm1*1s1_conv_add and gemm1*1s1_conv_add_bn_relu

* improve the performance of slidingwindow_bn_relu,slidingwindow_add,slidingwindow_add_bn_relu,gemm1*1s1_bn_relu,gemm1*1s1_add_relu

* prune faster depthwise

* fix style

* only InferShape when input shape has been changed

* apply input_dim_has_changed_ criterion only when model has one input vector

* add faster depthwise implementations

* fix style

* add test class to verify single op in a net

* fix private member error

* enhance scripts

* fix crash of tensor float * constructor

* fix fast check

* fix style