Possibly, depending on what you're going to use the CPMs for. If you're going to do something semi-quantitative (e.g., clustering on the log-CPMs), then it makes sense to incorporate the appropriate normalization information stored in y$offset. However, I usually just use cpm to get some values to plot, in which case the simplicity of the standard definition is more appealing. There is the extra complication that the offsets should be of roughly the same absolute scale as the log-library sizes, otherwise the output can't be interpreted as CPMs. This makes it difficult to guarantee that sensible values are returned by cpm in the presence of arbitrary user-defined offset matrices. (The last point isn't much a problem with the rest of edgeR, where the important thing is the relative size of the offsets between samples.)

Just to emphasize the last part of what Aaron has said. edgeR places no constraints on what users can input as the matrix of offsets. In fact one can add an additive constant to the offset values for each gene without changing any of the downstream results, except for the intercept coefficient estimate for each gene. This means that absolute sizes of the offsets have no meaning. Hence we cannot compute cpm (a quantity whose size has an absolute meaning) from offsets (whose size has no meaning).

I suppose we could go down a whole new development path of interpreting offsets in a stricter fashion, but no one has presented us with good reason why this would be worth spending time on.