geometric adstock optimization

[wd60622]

Code snippet from @aseyboldt which uses pytensor.scan to implement the geometric adstock. This doesn't require l_max

delay = pm.Gamma("delay", mu=5, sigma=2, dims="channel")

    def convolve_step(val, prev, delays):
        return val / delay + prev * (1 - 1 / delay)

    prev_zero = spending.T[0]
    
    convolved, _ = pytensor.scan(
        convolve_step,
        sequences=[spending.T],
        outputs_info=[dict(initial=spending.T[0], taps=[-1])],
        non_sequences=delay,
    )

Would have to be adapted to support arbitrary dimensions and map into GeometricAdstock class formulation