core: improve handling of local errors

Currently this (hacky) approach maps locally-generated error responses to error callbacks:
https://github.com/lyft/envoy-mobile/pull/457

Ideally, we'd create a generalized error pathway for filters in upstream Envoy, then use a filter to transform errors to local responses for the standard 'server' case and register error callbacks for the library case.

An intermediate change might be to provide a more deterministic signal on local/error responses to handle them more explicitly.

Both of these options would require upstream Envoy changes.