Two small changes to the TaskInput impls that compound across hundreds of
awaitee positions:

- Vc<T>::resolve_input now returns Vc::resolve() directly instead of an
  async fn body around to_resolved().await. The async fn envelope was
  pure overhead since the body had no other awaits — resolve() already
  returns a Future<Output = Result<Vc<T>>>, which is exactly the trait's
  signature. Using resolve() rather than to_resolved() + deref avoids
  the otherwise-discouraged-but-here-correct ResolvedVc round-trip.

- ResolvedVc<T>::resolve_input override removed. It was `async fn ... { Ok(*self) }`
  which is identical in size and behavior to the default
  `async { Ok(self.clone()) }` for Copy types.

Measured with -Zprint-type-sizes on next-napi-bindings:
- Per-Vc<X>::resolve_input future shrinks from 48 B to 32 B (-33%)
- Cascades into ~36 distinct concrete Vc<X> types and dozens of tuple
  TaskInput impls that contain Vc<T> fields
- Total awaitee-byte reduction across turbo/next contexts: ~94 KB