I have a dataset having 4 rna seq healthy tissue samples prepared with unstranded Illumina library and another dataset with 8 rna seq healthy tissue samples prepared with reverse stranded Illumina library. I wish to combine these 2 datasets together and run DGE with a third dataset having 100 tumor tissue samples. But DESEQ2 takes in raw counts as input. Can I normalize each dataset seaprately and then merge them together to run DESeq2?

Also, should I also run ComBat or sva after DESeq2 and then use that file as input to make PCA, t-SNE etc?

The vignette recommends to include any "correction" factors, be it categorical batch effect covariates or factors derived from the likes of svaseq/RUVSeq into the design. The general rule is that a dataset to be analysed in the analysis must be prepared together (in the lab) to avoid batch effects, and if multiple batches are present then each batch should contain samples of all groups. What you cannot do is take control from batch1 and tumor from batch2 and combine this, as this is called "fully confounded". As for PCA and other analysis the vignette recommends vst/rlog and you can regress batch effects (if not nested with experimental groups) with something like limma::removeBatchEffects(). Please be sure to read previous posts, this has been asked many times before and the support site is not intended to do hands-on guidance through your analysis.