Ensembl would be my choice as it will include things GenCode will not. Gencode curates on biological evidence for named features, while Ensembl will also include computationally predicted transcripts and genes. So you could think of it as Ensembl should let you cast a bigger net, to identify possible differences between your cell populations.

I have not used Salmon, but there are pre-built indices for BowTie2 available from Illumina if you want to skip building your own index - https://support.illumina.com/sequencing/sequencing_software/igenome.html

I have successfully used the Ensembl transcriptome with Salmon, but I think the Gencode set should be equally suitable.

If you decide to go for the Ensembl transcripts, you should be aware that the link you provide only contains transcripts coming from protein coding genes. If you are interested in also quantifying non-coding genes/transcripts, you should concatenate the cDNA file you provide with this file here

ftp://ftp.ensembl.org/pub/release-99/fasta/homo_sapiens/ncrna/Homo_sapiens.GRCh38.ncrna.fa.gz

before building the Salmon index.

(The Gencode file you indicate already contains non-coding transcripts.)

Oana is correct and for that reason I'd go for the Gencode set as it includes coding and non-coding transcripts. It has 227k transcripts vs 190k in ensembl.