Rather than drip-feeding an already-encoded response body through a
producer in 4KB chunks, hand the whole body to the transport at once and
only use a producer to delay finishing the request until the transport
has flushed it.
The transport only resumes a paused producer once its 64KB send buffer
is completely empty, so the old approach could send at most ~64KB per
reactor tick, which makes large responses crawl on a busy worker.
Matthew Hodgson
b7f9848429