I have 1,584 observations in my dataset, daily values for the last few years. What is a good rule-of-thumb for

input_size=XXX

, and what is

input_size

in lay terms as a parameter?

It is an empirical question, normally we guide the input_size as multiples of the seasonality of your dataset. For very long sequences I recommend WindowsBased models like NBEATS, NBEATSx, NHITS, MLP over RecurrentBased like LSTM, TCN, etc. Those models create temporal windows of size input_size+horizon.