fix(speculative): detect native MTP drafters via num_nextn_predict_layers (#2120)

resolve_drafter_kind fell back to dflash for a native checkpoint whose model_type
isn't in the kind table (e.g. deepseek_v4), and the dflash prefill path then read
config.target_layer_ids and raised before emitting a token. Auto-detect mtp when
the config declares num_nextn_predict_layers > 0, sharing one _expected_drafter_kind
helper with validate_drafter_compatibility.