You have it backwards. It's Turing who is the odd one out, as of the 128 CUDA Cores, Turing dedicated 64 for FP32 ops and 64 for INT32 ops.
Ampere reverted it back to all 128 CUDA Cores being able to do 128 FP32 ops, but with the ability of half the CUDA Cores being able to flexible do either FP32 ops or INT32 ops. Basically best of both worlds.
As for Maxwell, it's fully 128 FP32 ops, with INT32 being done by other means. So Ampere is very comparable to Maxwell in that sense.
And, anyway, regardless of whatever operation the CUDA cores are doing, T239 GPU still has 1536 Shading Units, which is 6x the Tegra X1/T210(4), which had 256 Shading Units. As Kenka complemented, a 6x bigger GPU + others advancements means it will be fairly easy for Switch 2 to be a >6x jump in GPU performance.