[Feature] Support Speculative Decoding (MTP)

[wmertens]

Summary

Gemma and other models allow for speculative decoding.

Problem / Motivation

Generation is slow.

Proposed Solution

https://blog.google/innovation-and-ai/technology/developers-tools/multi-token-prediction-gemma-4/

Alternatives Considered

Platform

• Android
• iOS
• Both

Additional Context

[CompactAIOfficial]

fyi Speculative Decoding (SD), Speculative Speculative Decoding (SSD) are separate from MTP
but would be nice for MTP to be added