Gemma 4 assistant-head speculative decoding (MTP)

[Michael Burch (michaelburch)]

Add support for MTP with Gemma 4.

Create a branch that adds assistant-head speculative decoding for Gemma 4, reusing the existing MTP path. Tested end-to-end, lossless vs greedy. 4 commits:

Extract shared MTP generator base — refactor, no behavior change. Splits MtpGenerator into MtpGeneratorInterface + MtpGeneratorBase so a second generator can share the plumbing. Independently useful.
Add Gemma 4 assistant-head speculative decoding — new Gemma4AssistantGenerator, config schema (max_logits_sequence_length, shared_kv_layers, main_inputs_embeds), plus logits/hidden_states/multi_modal plumbing.
Export script + docs — python.
Simplify Logits::Update — complexity cleanup on code touched by (2).

Questions:

Export path. builder.py has no Gemma 4 support — it stops at Gemma3ForConditionalGeneration. My export lives in python and uses an external exporter. Should it go in builder.py instead?

Direction. README lists speculative decoding as roadmap. Is an architecture-specific assistant head the intended shape, or is a general mechanism planned?

Config schema. The three new genai_config.json fields are public surface — naming and placement acceptable?

Shared code. Changes to logits/hidden_states/multi_modal affect all models. Prefer those split into their own PR?

Happy to land (1) alone first if the rest needs iteration.