An ideal delay and sum beamformer is Linear, which implies that filtering before beam forming is equivalent to filtering after beam-forming. If you filter each sensor, the filters have to be matched very closely, so for a pure analog system, do you want to use one or N nearly identical filters. Given temperature sensitivity and parameter drift, filtering after is much cheaper. No implemented beam former is perfectly Linear. Filtering prior has a number of advantages, such as reducing dynamic range requirements. Also for a digital beamformer, you need an antialiasing filter prior to the A/D converter so you filter before anyway. 30 years ago, switched capacitor filters were very useful. Today, for audio, over sampled delta sigma converters are not expensive and have huge dynamic ranges but you still have to do some signal conditioning between the mechanical sensor and the converter. The trade off for RADAR is more complicated. If your converters have enough dynamic range.
So the answer is that filtering prior to beamforming needs nearly identical characteristics. Some deviation is ok but you need to make that determination for your self . Ceramic and moving coils are not ideally Linear phase devices either and sensor location uncertainty has an effect. You also need to consider that you might have nonacoustic paths between the source and your array, Element coupling is another problem that can emerge.