feat(constants): wire up backend selection in MODEL_REGISTRY accessors

Each accessor's `backend` parameter is now typed to exactly the backends the
model ships with — passing an unsupported one is a compile-time error.
`Platform.OS` still picks the default when `backend` is omitted. The per-
backend (quant × backend) variant matrix lives in modelRegistry.ts so
modelUrls.ts stays flat-per-model.

Unifies DISTILUSE_BASE_MULTILINGUAL_CASED_V2 to one accessor with
xnnpack + coreml; the _8DA4W and _COREML named constants stay as
deprecated aliases.

Author: msluszniak

docs/docs/05-utilities/model-registry.md

@@ -36,14 +36,14 @@ const llmQuant = useLLM(MODEL_REGISTRY.LLM.LLAMA3_2_3B({ quant: true }));
 ## Options
 
 ```typescript
-type ModelOpts = {
+type ModelOpts<B extends Backend = Backend> = {
   quant?: boolean; // Pick the quantized variant when true. Ignored when no quantized variant exists.
-  backend?: 'xnnpack' | 'coreml' | 'vulkan' | 'qnn'; // Reserved; today the platform default applies.
+  backend?: B; // Explicit backend; the set of allowed values is per-model.
 };
 ```
 
 - `quant: true` resolves to the quantized variant for models that publish one (e.g. `LLAMA3_2_3B` → SpinQuant, `EFFICIENTNET_V2_S` → int8). For models with a single variant, `quant` is accepted but has no effect.
-- `backend` is accepted in the signature for forward-compatibility. Today every accessor returns a config whose backend is selected at module-load time by `Platform.OS`. Explicit per-backend selection is planned for a future release.
+- `backend` selects an explicit backend. The accessor's call signature is typed to exactly the backends the model ships with, so requesting one the model doesn't publish is a compile-time error (e.g. `MODEL_REGISTRY.LLM.LLAMA3_2_3B({ backend: 'coreml' })` does not type-check — Llama 3.2 is xnnpack-only). When `backend` is omitted, the platform-default applies (typically CoreML on iOS and XNNPACK on Android for multi-backend models).
 
 ## Usage patterns
 
@@ -70,6 +70,20 @@ const classifier = useClassification(
 );
 ```
 
+### Explicit backend
+
+```typescript
+// Force XNNPACK on iOS (overrides the CoreML default).
+const detector = useObjectDetection({
+  model: MODEL_REGISTRY.OBJECT_DETECTION.RF_DETR_NANO({ backend: 'xnnpack' }),
+});
+
+// Combine `quant` and `backend`.
+const styled = useStyleTransfer(
+  MODEL_REGISTRY.STYLE_TRANSFER.CANDY({ backend: 'coreml', quant: true })
+);
+```
+
 ### Inspecting a model's config
 
 Accessors transparently expose the default config's fields, so you can read them without calling: