power_t: does it make sense for this parameter to have negative values

[reshamas]

sklearn/linear_model/_stochastic_gradient.py

Ref #22115

I do not anticipate negative power_t to be mathematically meaningful but apparently our code accepts it without crashing... So ok with documenting it.

Originally posted by @ogrisel in #22115 (comment)

@thomasjpfan

I feel like this a case where documenting -inf will lead to more people trying out. If this is not mathematically meaningful, then we could be promoting a bad practice?

cc: @glemaitre

[urbanophile]

My $0.02 is that this should be restricted to (0, inf). I have only heard of people wanting to increase the learning rate when doing something like cosine annealing, a linear warmup and decay or a cyclic learning rate for much more complicated deep neural networks. In the context of a linear classifier, I think you want the whole method to be pretty well-behaved, and the algorithm will not converge if the learning rate increases.

So, I'm happy to change this and submit a pull request if other people are agreeable.