I found some interesting dedicated hardware that could help, namely the AD9833 but more specifically the AD5932, which has a programmable acceleration ramp.
In the mean time I think I'll just go with a "loop-based" acceleration and deceleration function (hardcoded delayMicroseconds with varying acceleration factor) and swith to straight hardware PWM in between. That way my CPU will be available except when changing speed, which takes about 500ms in my case...