`next-dev-server` having its own implementations of `renderToHTML` and `renderErrorToHTML` has historically made reasoning about streaming hard, as it adds additional places where status codes are explicitly set and the full HTML is blocked on.
Instead, this PR simplifies things considerably by moving the majority of the custom logic for e.g. hot reloading and on-demand compilation to when we're resolving the page to be loaded, rather than upfront when handling the request. It also cleans up a few other details (e.g. default error page rendering) that managed to creep into the base implementation over time.
One unfortunate side effect is that this makes compilation errors slightly more difficult. Previously, we'd render them directly. Now, we have to rethrow them. But since they've already been logged (by the watch pipeline), we have to make sure they don't get logged again.
9ab916ac99