Wonder why they wouldn't pass in the whole request on stdin? That would also mean that instances could be pre-started in anticipation of a request and not just when one is already waiting.
As for the fork() per request, I believe mostly because they were starting with requirements where avoiding fork() wasn't important. Conserving memory was probably more important, given the timeframe. CGI programs that leaked memory would have been common if they were daemons. For most web servers, there also probably weren't that many requests.
It was also a common existing pattern. Many of the services on unix boxes worked the same way...one daemon (inetd) would start instances of daemons per request and they would serve only one request.
That said, NSAPI appeared pretty quickly after CGI, then FastCGI, etc. Still pushing environment variables for headers, but stopping the request->fork->exit pattern and allowing persistent daemons.