Add Gemma 4 MTP scaffolding

[ncylich]

Summary

• add Gemma 4 MTP assistant loading, speculative draft/verify flow, sampling support, and metrics
• add KV cache transaction support and focused MTP/unit coverage
• keep MTP disabled by default with explicit draft-count opt-in

Validation

• source ./venv/bin/activate && cactus build && ./cactus-engine/test.sh
• CACTUS_TEST_GEMMA4_TARGET=/Users/noahcylich/Documents/Desert/cactus-v2/weights/gemma-4-e2b-it CACTUS_TEST_GEMMA4_ASSISTANT=/Users/noahcylich/Documents/Desert/cactus-v2/weights/gemma-4-e2b-it/assistant ./tests/mtp/run_fast_gemma4_mtp.sh --max-tokens 24 --timeout-seconds 240
• ./tests/bench_e2e_decode.sh --model gemma-4-e2b-it --prompt "Count from 1 to 100, separated by commas." --max-tokens 80 --reps 1 --warmup 0 --mtp-max-drafts 1,2,3 --output /private/tmp/mtp_1_2_3_pr_verify.csv

MTP sweep

• draft 1: 13.99 decode tok/s, 100.0% accept, no fallback
• draft 2: 19.48 decode tok/s, 98.1% accept, no fallback
• draft 3: 23.21 decode tok/s, 96.7% accept, no fallback