With the recent change to stream the request body to the backend like
chunked encoding does, even when we know the content length, two test
cases became sensitive to the timing of window updates. The data frames
contain for the most part a 16kB payload, which happens to be the same
as fetch_chunksize's default value.

With a value of 64kB, corresponding to one more byte than the default h2
initial window, we are garanteed to never risk a race between the VFP
consuming the h2_rxbuf and the h2_sess task monitoring it to send window
updates.

This is also a good occasion to exercise stream.peer_window in t02019.

Refs #3878
Refs #3934