A non-rectangular window will remove "noise" from distant bins at the cost of adding more "noise" to the immediately adjacent bins to a narrow-band spectrum peak. The sum of both these spectral leakage effects is greater than zero for a non-rectangular window. So if you count the raising of the level total of all adjacent bins as noise, then the S/N ratio is lowered.
Some people don't care about the bins immediately adjacent to a spectrum peak (their spectral peaks are a priori assumed to be widely spaced; and/or they interpret, interpolate, or phase-vocoder adjust the energy out of those adjacent bins back into the central peak bin), so for those purposes, the reduced far-side-lobe energy means less noise.
Another reason for a lower S/N ratio is that windowing of quantized data is an informationally lossy process, and these (re)quantization losses can also be considered a form of noise.
"Best" is relative to some weighting of quality metrics, and different users may have very different weightings. If you don't have a set of prioritized design goals for which to optimize a window, then you may not have a strong reason to not just use a Von Hann window.
Depending on your data source and your needs, using just some windowed FFTs may not even be a good form of spectrum analysis, much less the best possible. Or the opposite.