Expected behavior for repeated `fit!` calls

Just wondering what the right strategy is for warm starts. What is the correct behavior for calling fit! on a machine twice? Here, I assume that the model has an $N$ parameter controlling the number of optimization steps:

model = Regressor(N=N)
mach = machine(model, X, y)
fit!(mach)
fit!(mach)  # what happens here?

1. Warm start; runs for another $N$ steps from where it left off.
2. Warm start; runs for $0$ steps from where it left off. In other words, the user would need to increase $N$ if they want to replicate the behavior of (1)

   fit!(mach)
   mach.model.N += N
   fit!(mach)
   

3. Cold start; resets the search state and runs for $N$ steps. Perhaps you need to use an update! function for explicit warm starts (in which case 1 and 2 would require using update! instead)

Right now, SymbolicRegression.jl is doing 2. But I'm not sure about this. Pinging @ablaom for tips.

Start by examining the fit! entry point and the existing repeated-call behavior described for SymbolicRegression.jl. Determine and document the intended warm-start contract for a second fit!, including how the N parameter should behave; no specific files or tests are named in the issue.