IMHO it's not all that an astute question, insofar as upsampling during warping is commonly used and is hardly controversial. (There's lots of pitfalls and problems in fMRI analysis; this simply isn't one of them.)
One minor quibble: when you write, "therefore the number of contrasts," you presumably mean the number of statistical tests. The number of contrasts, identified by a set of contrast weights, remains the same; the number of tests goes up since there's one test per voxel per contrast.
It's true that upsampling leads to more statistical tests, which notionally worsens the multiple comparison problem. However, the MCP is already pretty awful for neuroimaging---the worsening of the MCP in your example is dwarfed by the base size of the MCP even with the original voxel sizes.
Re your specific question, upsampling isn't really relevant in terms of the practical steps one takes to try to ameliorate the MCP, or in terms of the cautions one must be mindful of when interpreting activation patterns in the absense of rigorous correction. As you point out, the "best" correction would use the "number of independent tests", which one would think would remain essentially unaffected by upsampling.