Can't background a job after a previously-backgrounded job completes

[ridiculousfish]

This tracks the following regression relative to 3.1.2, I'm opening this so it doesn't get dropped:

1. Run sleep 5
2. Hit control-Z to background it
3. fg to foreground it
4. Allow the sleep complete normally
5. Run sleep 100 and try to hit control-Z to background it

it will no longer background. stty -a shows the tty mode goes from susp = ^Z (correct) to susp = <undef> (the bug).

This bisects to 3d2dc85 however that fix only revealed a latent bug; the root cause is the fix for #2114.

[ridiculousfish]

Notes to myself mostly but fun reading if you are inclined:

The issue is annoyingly convoluted. Interactive fish stores two tty modes: shell_modes for itself, and tty_modes_for_external_cmds for external procs. For example, ctrl-Z should send SIGTSTP for external procs, but not for fish, because maybe you want to bind ctrl-Z (per 3d2dc85).

Now maybe your external process is literally stty which changes the tty modes. It wouldn't be very useful if those changes were dropped, so when fish gets control back it reads the tty mode and updates tty_modes_for_external_cmds.

The fix for #2114 caused us to more eagerly restore the tty to shell-mode when a previously-backgrounded job completes. So we restore the tty to shell-mode, then later we save that as the mode for external procs. So now every proc is like a shell, which is why ctrl-Z fails.

The desire in #2114 was to support foreground and backgrounding jobs from key bindings (!), not from ordinary commands. The proper fix is to extend the tty-mode save/restore to jobs spawned from key bindings, not just those typed at the command line.