refactor: remove dead code, deduplicate backends

Phase 1 - delete dead code:
- studio.py, migrate_add_instruct.py, utils/validation.py
- duplicate _profile_to_response in main.py, duplicate asyncio import
- pointless _get_profiles_dir/_get_generations_dir wrappers
- duplicate LANGUAGE_CODE_TO_NAME and WHISPER_HF_REPOS constants

Phase 2 - extract backends/base.py with shared utilities:
- is_model_cached() replaces 7 copy-pasted HF cache checks
- get_torch_device() replaces 5 device detection methods
- combine_voice_prompts() replaces 5 identical implementations
- model_load_progress() ctx manager replaces progress boilerplate in all backends
- patch_chatterbox_f32() replaces identical monkey-patches in both chatterbox backends

net -1078 lines across the backend

backend/export_import.py

@@ -19,11 +19,6 @@ from .models import VoiceProfileCreate
 from . import config
 
 
-def _get_profiles_dir() -> Path:
-    """Get profiles directory from config."""
-    return config.get_profiles_dir()
-
-
 def _get_unique_profile_name(name: str, db: Session) -> str:
     """
     Get a unique profile name by appending a number if needed.
@@ -99,7 +94,7 @@ def export_profile_to_zip(profile_id: str, db: Session) -> bytes:
 
         # Create samples.json mapping
         samples_data = {}
-        profile_dir = _get_profiles_dir() / profile_id
+        profile_dir = config.get_profiles_dir() / profile_id
 
         for sample in samples:
             # Get filename from audio_path (should be {sample_id}.wav)
@@ -181,7 +176,7 @@ async def import_profile_from_zip(file_bytes: bytes, db: Session) -> VoiceProfil
             profile = await create_profile(profile_create, db)
 
             # Extract and add samples
-            profile_dir = _get_profiles_dir() / profile.id
+            profile_dir = config.get_profiles_dir() / profile.id
             profile_dir.mkdir(parents=True, exist_ok=True)
 
             # Handle avatar if present