1. enable conv mps with texture bias. 2. add scale mps. 3. remove duplicate computation in fetch kernel. 4. fix typo. (#1665)

* 1.enable both tensor and texture as bias of conv_add_relu. 2. add conv_relu fusion op.

* winograd could be used when. out channel is less than or equal to 16

1. add optimize parameter to load function. 2. add transpose operation if the input of fetch op is not [0, 2, 3, 1]. (#1651)

1. reshape output transpose should be 0, 1, 2, 3. 2.mps conv should be disabled when groups is larger than 1. 3. slice should re-tranpose its input if needed. (#1646)

1.add groups to conv3x3. 2.add original size to tensor and tenture and adjust concat layer to handle tensor truncation logic. (#1643)

1.add flatten2, slice and nearest_interp op. 2.truncate tensor with more than 4 dimensions to 4. (#1634)

1.parse tensor desc info from params file. 2.ensure tensor desc is consistent between params file and model file. (#1631)

* shift filter conversion process to loading phase

* fix winograd edge case

* implement winograd for metal depthwise 3x3 naively

* keep half4 in non-winograd depthwise 3x3 to ensure precision

* fuse Conv-Add-Relu into one op.

* Revert "fuse Conv-Add-Relu into one op."

This reverts commit 200be579.

* fuse Conv-Add-Relu into one op