Add LM Backend selector (PT vs VLLM) to reduce VRAM usage

PT backend uses ~1.6 GB VRAM vs ~9.2 GB for VLLM, making it accessible
on lower-end GPUs. Adds dropdown in Advanced Settings defaulting to PT.

services/api.ts

@@ -232,6 +232,7 @@ export interface GenerationParams {
   lmTopK?: number;
   lmTopP?: number;
   lmNegativePrompt?: string;
+  lmBackend?: 'pt' | 'vllm';
 
   // Expert Parameters
   referenceAudioUrl?: string;