Sharing Encoder Representations across Languages, Domains and Tasks in Large-Scale Spoken Language Understanding