llama.cpp Reflects MTP Improvements

llama.cpp's MTP clean-up PR #23269 was merged into master on May 19.

The key fixes are as follows.

• Fixed a parallel decoding bug with recursive rollback (n_rs_seq > 0).
• --spec-draft-p-min works again, with its default reverted to 0.0.
• Lowered the --spec-draft-n-max default from 16 to 3.
• Changed all speculative implementations to share accepted tokens, improving acceptance rate when chaining multiple speculative methods.
• Fixed ngram-k4v parameter passing and a batch.token && batch.embd graph reuse issue.

The author explained that parallel decoding with MTP is not yet well optimized and needs a larger refactor. Instead, the startup log now prints settings for each speculative type, and an example of using multiple spec methods together has been added.