server: implement GLM-style MTP#15225

Draft

F1LM1 wants to merge 11 commits intoggml-org:masterfrom

F1LM1:glm4-moe-mtp

Commits on Aug 11, 2025

Commits on Aug 12, 2025

Commits on Aug 13, 2025

broad thrust of the mtp implementation
F1LM1
committed

Commits on Aug 16, 2025

failed attempt to implement MTP; outputs tokens but KV cache management is unreasonable
F1LM1
committed

Commits on Aug 17, 2025

added proper KV cache management for MTP layers and slightly refactored
F1LM1
committed

Commits on Aug 18, 2025

fixed mtp kv cache update sequencing after prompt processing
F1LM1
committed

Commits on Aug 19, 2025

kludge-y kv cache management of mtp layer
F1LM1
committed

Commits on Aug 20, 2025

fixed vram leak
F1LM1
committed

Commits on Aug 26, 2025

replace standard sampler with greedy sampler for mtp draft
F1LM1
committed