xlat/gen.sh: speedup xlat generation

This commit implements the following changes:
- increases the number of concurrent jobs to twice the CPU count;
- creates a circular buffer, so instead of running multiple jobs
  at once, the generator tries to keep about the same number of jobs
  being run concurrently;
- runs gen_git and gen_make concurrently in order to squeeze in
  one more bit of concurrency.

With my deeply scientific measurements, this approach achieves
up to 15% speedup with dash and about 30-40% with bash as /bin/sh
on a 4-core machine.

* xlat/gen.sh (main): Declare pids local variable, append pid of every
run job to it; increase the limit of jobs to ncpus * 2; when the limit
is reached, wait for the first pid in pids instead of resetting jobs
to zero and waiting for all the run jobs; put gen_git and gen_make
into background.

xlat/gen.sh

@@ -290,6 +290,7 @@ main()
 	local name
 	local jobs=0
 	local ncpus="$(getconf _NPROCESSORS_ONLN)"
+	local pids=
 	[ "${ncpus}" -ge 1 ] ||
 		ncpus=1
 
@@ -300,15 +301,20 @@ main()
 			name=${f##*/}
 			name=${name%.in}
 			gen_header "${f}" "${output}/${name}.h" "${name}" &
+			pids="$pids $!"
 			names="${names} ${name}"
 			: $(( jobs += 1 ))
-			if [ ${jobs} -ge ${ncpus} ]; then
-				jobs=0
-				wait
-			fi
+			if [ "${jobs}" -gt "$(( ncpus * 2 ))" ]; then
+				read wait_pid rest
+				pids="$rest"
+				wait -n 2>/dev/null || wait "$wait_pid"
+				: $(( jobs -= 1 ))
+			fi <<- EOF
+			$pids
+			EOF
 		done
-		gen_git "${output}/.gitignore" ${names}
-		gen_make "${output}/Makemodule.am" ${names}
+		gen_git "${output}/.gitignore" ${names} &
+		gen_make "${output}/Makemodule.am" ${names} &
 		wait
 	else
 		name=${input##*/}