Add Support for Xing4.0 (cleanup)

[fairydreaming]

Overview

This is an experimental PR where I'm trying to clean up #29012 by simplifying/removing duplicate code.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[fairydreaming]

@ggerganov @am17an I managed to reuse DSv4 mHC OPs in Xing4.0 impl while maintaining bit-exact results, but I had to add another variant of ggml_dsv4_hc_comb() - ggml_dsv4_hc_comb_clamp() with added float limit argument (and one additional ggml_transpose() call in the model implementation).

Internally the clamp variant works a little different from the original. In DeepSeek V4 positivity of the comb matrix before Sinkhorn transform was enforced by performing softmax and adding eps to all matrix elements, while in the variant with clamping positivity of the matrix is enforced by the choice of the limit value (so that exp() of clamped values - max is non-zero when expressed in float) then the Sinkhorn loop can perform all norm iterations with eps added only in the denominator.

I wanted to know if this solution is acceptable or you have some other idea before I mark this ready for full review.

[am17an]

if you introduce new params you need to also take care of other backend which support this op to return false in supports_op. In terms of the actual impl for this model it looks ok to me to do this.