Created by: FrostML

Fix seqconv_eltadd_relu fault when using cpu multi-thread inference

Before this PR, the seqconv_eltadd_relu_fuse_pass will randomly or occasionally cause a seg-fault when using cpu multi-thread inference with cloned predictors.

This PR fix this bug by removing the "ColMat" in AddOuput().AsIntermediate() and provide a intermediate tensor during kernel computing.