feat: speculative decoding support

[jundot]

summary

adds scheduler-level speculative decoding for faster single-request generation on apple silicon.

• when only 1 request is in the batch and a draft model is configured, the scheduler drafts N tokens with the small model then verifies all N+1 in a single main model forward pass
• automatically falls back to standard continuous batching when multiple concurrent requests are active
• supports hybrid SSM/attention models (Qwen3-Next etc) via cache snapshot/restore instead of trim
• fixes simple KVCache trim bug where offset resets on next update_and_fetch
• admin dashboard UI for configuring draft model and num_draft_tokens per model
• per-request cumulative accept rate logging

files changed

area	files
core	scheduler.py (speculative step, draft cache management, snapshot/restore)
engine	engine/batched.py, engine/vlm.py, engine_pool.py (draft model loading)
settings	model_settings.py (new fields), model_discovery.py
admin	routes.py, dashboard.js, modal/settings templates, i18n (ko/en/ja/zh)
tests	test_model_settings.py

test plan

• single LLM request with speculative decoding enabled — verify accept rate in logs
• multi-request scenario — verify automatic fallback to standard batching
• hybrid SSM model (Qwen3-Next) — verify snapshot/restore produces correct output
• pure attention model — verify trim-based cache rewind works
• admin UI — toggle speculative decoding, select draft model, save/reload settings
• VLM model with draft model configured — verify it loads and works
• pytest tests/test_model_settings.py -v