docs(speculative): document MTP, DFlash, and DSpark support

- move speculative decoding next to sampling configuration in the README
- document stateful MTP, DFlash, DSpark, and n-gram engine configuration
- add DFlash and DSpark external-sidecar examples and tuning guidance
- describe block-size limits, backend sampling, checkpoints, and statistics
- document the current text-only and single-sequence limitations
- record tested Qwen3.5, Qwen3.6, and Qwen3.8 MTP configurations
- document tested external MTP support for gemma4 with gemma4-assistant
- explain automatic target-context, shared-KV, and same-position handling
- link the dedicated MTP and DFlash/DSpark benchmark examples

Signed-off-by: JamePeng <jame_peng@sina.com>

Author: JamePeng