You already have the lowpass filter that you need. It's the control loop round the beam you are positioning.
You only need do as much lowpass filtering of the sensor that your control loop does automatically. For instance, if you're running a digital loop with a control repetition period of 1mS, then average all the sensor readings in 1mS and present them to the control loop. This will have no more phase shift than the latency inherent in the loop. If an analogue loop, then just drive the existing electronics with the wideband sensor signal. You'll have designed some sort of lowpass filter into the system and taken account of it when considering stability.
You are correct that any additional filtering after the sensor will eat into your stability margin, which will probably need a reduction in loop bandwidth to regain stability.
As you strive to make the loop wider, you will follow the sensor to higher modulation frequency. If the sensor is inherently noisy, you will follow more of the noise. Conventionally what we do is to choose a loop bandwidth as a compromise between the sensor and the effector noise, to get minimum overall output contribution from both.
If you have a noisy sensor, then your two simple options are
a) get a quieter sensor
b) reduce the loop bandwidth
There may be cleverer things you can do, depending on why you are striving for a wide loop bandwidth. If it's for rapid dynamic motion control, then you're stuck with the sensor noise. If OTOH it's to keep the beam on station, statically, then you could split the sensor into two types. You could use your existing noisy DC coupled sensor to control the mean position, locked in a narrow loop. You could then use an AC-coupled lower noise sensor, say a large test mass, magnet and pickup coil, that is a velocity sensor, to control a wider bandwidth 'quietening' loop. It depends on the specification for your dynamics whether this could be used. You could for instance use switchable loop bandwidths to get rapid noisy slew to a new position, then narrow the position loop and enable the quietening loop, though this does produce a short period of time when the output is not fully controlled for both position and noise. It would also need a fair bit of understanding to implement it. I've used the technique for low noise RF synthesis, when there is a permitted settling time, so it works. Think of it as synthesising a lower noise sensor.