The function fft allows one to zero-pad, and mscohere, after all works with auto and cross spectra, which, presumably, have all been calculated with fft. I haven't found any mention on zero-padding in the documentation for mscohere.
have you tried >> type mscohere.m ? It looks like you need to do is modify the call to welch.
The routine doesn't calculate a confidence region so, yes you could formulate a call to a coherence calculation with zero padding but the sample probability distributions published in any of the papers by Cliff Carter, Al Nuttal, and Charlie Knapp would not apply. The biggest error with using mscohere is not averaging enough. Each bin needs around 500 independent samples to have roughly a $\pm$10% accuracy of the full value range of $[0,1]$. Given the independent averaging issue, and that zero padding does nothing for that, it's hard to see a justification.
SNR effects the value of the coherence, not the accuracy of the estimate. It is not a technique suitable to small sample sets.