Fix several bugs in GRPC stream handlers and state machine

[gjcairo]

This PR fixes three bugs introduced in recent changes:

• We weren't correctly handling the (invalid) scenario in which a server may send a data frame with EOS (instead of closing by sending status + trailers). This is now handled correctly by forwarding an internalError status code down the pipeline, and closing both client and server.
• We were forwarding channel close when the close mode was output. However, because the H2 stream handler ignores the close mode and just closes down the whole channel upon receiving close, this was inadvertently closing the whole stream. We now only forward the close if the mode's all.
• We were retrieving the grpc-status-message header using the canonical form, which caused us to only retrieve part of the original message if it included a comma. This has been fixed.

[glbrntt]

We're leaking a promise but looks good otherwise