Also add `Partial` associatedtype to `Generable` protocol with default Self.

All implementations currently throw "not implemented" errors.

`@Generable` now emits a nested subtype for streaming use cases:

```
@generable
struct T {
  // Fields here.
}
```

now expands to:

```
struct T {
  // Fields here.
}

extension T: Generable {
   Partial {
      // partial fields
   }

  // Other generated fields.
}
```

The rules for generating the partial subtypes are as follows:

- `T?` => `T.Partial`,
-  otherwise: `T` => `T.Partial`

This should help avoid unwanted automatic defaults

Currently none of the backends support streaming, so marking the tests are disabled.

Tool use is still not supported

When a message is truncated, OpenAI returns `incomplete` event rather than `completed` event.

To avoid code duplication

- Unify generateResponse and generateResponseStream implementations
- Both methods now use the same underlying streaming mechanism
- Reduce code duplication and improve consistency
- Maintain same API contract and error handling

Removed the 'sending' keyword from all AsyncThrowingStream return types
across the streaming API. This affects:

- LLM protocol methods (replyStream functions)
- All backend implementations (OpenAI, Apple, MLX)
- Test implementations (FakeLLM, CrashingLLM)
- Documentation examples

The streaming functionality remains unchanged, only the function
signatures have been updated to remove the sending parameter.