feat: add experimental MTP setting, off by default

[kdroidFilter]

Summary

• Add a Settings toggle for LiteRT-LM Gemma 4 multi-token prediction (MTP), labelled experimental and off by default.
• When enabled, set ExperimentalFlags.enableSpeculativeDecoding before engine creation (Android, desktop, and the Linux GPU worker).
• Persist the flag in the settings snapshot; older installs stay off. Reloading the model picks up a change (immediately if keep-alive is on).

Test plan

• :shared:compileKotlinJvm / :shared:compileAndroidMain
• :shared:jvmTest (AppViewModelTest, StringResourcesTest, LinuxGpuPolicyTest)
• Open Settings → Model and confirm Multi-token prediction (experimental) is Off
• Turn it On, translate on GPU, confirm the engine reloads and decode is faster