Fix the order of BE kern gathering

[rsheeter]

I believe this resolves the issue described at #647 (comment) by applying the same fix as #655 to the second scatter/gather we do on kerning.

I have filed #677 to track making this simpler as it's definitely harder than it shoudl be.

The following produces observably unstable output on main, usually in 2-4 builds, and is stable through 20 cycles after the fix:

$ rm -f *.ttxl threads*.svg && for i in {0..19}; do echo "CYCLE $i" && rm -rf build/ && cargo run ../googlesans/source/GoogleSans/GoogleSans.designspace --emit-timing && cp build/threads.svg ./threads_$i.svg && ttx -l build/font.ttf > run$i.ttxl && grep GPOS *.ttxl; done

# Unstable: GPOS occasionally dramatically smaller
# because gather ran before the tasks that produce what it gathers
CYCLE 2
    Finished dev [unoptimized + debuginfo] target(s) in 0.19s
     Running `target/debug/fontc ../googlesans/source/GoogleSans/GoogleSans.designspace --emit-timing`
run0.ttxl:    GPOS  0x19131F6A   4364690     19000
run1.ttxl:    GPOS  0x19131F6A   4364690     19000
run2.ttxl:    GPOS  0x4F3B316D    106928     15264  # MUCH BADNESS

JMM if happy.

[cmyr]

if you're happy, I'm happy.