fg: resume_job runs before SIGCONT is confirmed (same ordering bug as bg, GH #126 part B)

[tobert]

Follow-up from #126 part B (#160).

fg.rs (lines ~107-112) has the textually identical ordering bug that was
just fixed in bg.rs: manager.resume_job(job_id).await (clearing the job's
Stopped flag, marking it JobStatus::Running) runs before
killpg(pgid, SIGCONT). If the signal fails (e.g. the process already died),
the job is left looking Running with no live process.

// fg.rs, current
if let Err(e) = term.give_terminal_to(pgid) {
    return ExecResult::failure(1, format!("fg: failed to give terminal: {}", e));
}

// Mark as resumed and send SIGCONT
manager.resume_job(job_id).await;
if let Err(e) = nix::sys::signal::killpg(pgid, nix::sys::signal::Signal::SIGCONT) {
    let _ = term.reclaim_terminal();
    return ExecResult::failure(1, format!("fg: failed to continue job: {}", e));
}

Deliberately left out of #160's scope (which was bg.rs only, per #126's own
wording) because fg has more surface to get right than a mechanical
reorder:

• term.give_terminal_to(pgid) also runs before killpg — handing the
  terminal to a process group that's about to fail to resume (or is already
  dead) is arguably its own instance of the same bug. A correct fix likely
  needs to confirm SIGCONT before both resume_job() and the terminal
  handoff, with reclaim_terminal() called on every failure path (it already
  is on the SIGCONT failure branch, but the ordering needs to be rethought
  end to end).
• Unlike bg, fg cannot be unit-tested without a real controlling terminal
  (TerminalState::init() calls real tcsetpgrp/setpgid and has no
  test-friendly constructor) — a regression test for this will need to be
  PTY-based, extending crates/kaish-repl/tests/pty_job_control.rs.

Flagged during kaibo review of #160 (cast: deepseek), which agreed the split
was reasonable given the added complexity here.