`base_score` is redundant for multinomial logistic

[david-cortes]

After this PR: #10298

Objectives multi:softprob and multi:softmax (which correspond to multinomial logistic) calculate an intercept based on averaging values across all classes.

But this has no effect since adding a constant base score to all columns does not change the result after applying a softmax transformations - that is: softmax(a, b, c...) == softmax(a + const, b + const, c + const, ...).

It could skip the calculation altogether in such case.

[trivialfis]

I will move ahead and remove the binary model in this release: #11307

Afterward, we can have a vector for the base score.

[trivialfis]

It could skip the calculation altogether in such case.

We still need the intercept as a base model for consistency. But you have a very good point. I can relax some constraints.

[david-cortes]

It could skip the calculation altogether in such case.

We still need the intercept as a base model for consistency. But you have a very good point. I can relax some constraints.

But you could leave it fixed at zero then. It wouldn't change the end result.

[trivialfis]

We are moving toward vector-valued intercept. #11656