Make multiclass base_score consistently use response-space values

Multiclass handles base_score differently from other objectives with an invertible link.

For logistic, Gamma, Poisson, and Tweedie objectives:

1. InitEstimation returns a response-space value.
2. ProbToMargin converts it to the internal margin.
3. A user-provided base_score is interpreted in response space.

Multiclass instead converts estimated class probabilities into centered logits directly inside InitEstimation and does not implement ProbToMargin. Consequently, a user-provided vector base_score is interpreted as logits rather than class probabilities.

Multiclass has a suitable canonical conversion despite softmax being invariant to a constant:

margin[i] = log(p[i]) - mean(log(p))

Proposed changes:

- Return class probabilities from multiclass InitEstimation.
- Implement ProbToMargin using clamped, centered log probabilities.
- Add tests comparing response-space base_score with the equivalent raw base_margin.
- Handle compatibility with existing models, which may store vector-valued multiclass base_score as already-transformed margins.

This would make multiclass consistent with the documented base_score convention while retaining a deterministic representative of the equivalent softmax margins.

Start by locating the multiclass InitEstimation implementation and the objective interface for ProbToMargin. Review existing base_score and base_margin tests, then trace how older vector-valued multiclass models are loaded. Done means response-space base_score matches the equivalent raw base_margin, with compatibility tests covering existing models.

Written by the indexing model from the issue text.