Fix pipe hang (issue #319)

Turns out the kernel needs to handle blocking file descriptors differently, it
needs to do thread-pooling and it's a different code path (also slower).

It also seems that this is bugged in some cases. This makes sense as I believe
most people don't test with blocking FDs. I believe we should make all our FDs
non blocking as a precaution.

There is a program listed in issue #319 that replicates this. I could hang it in
< 5 seconds, and it's now running for 10minutes with the fix, so kaboom.

A fix has been committed to mainline Linux (should be in Linux 6.1):
torvalds/linux@46a525e

lib_eio_linux/eio_linux.ml

@@ -1290,6 +1290,8 @@ let stdenv ~run_event_loop =
 
 let pipe sw =
   let r, w = Unix.pipe () in
+  Unix.set_nonblock r;
+  Unix.set_nonblock w;
   let r = source (FD.of_unix ~sw ~seekable:false ~close_unix:true r) in
   let w = sink (FD.of_unix ~sw ~seekable:false ~close_unix:true w) in
   r, w