基准测试测出来几乎为零

// A: result thrown away, one sync at the end.
let t = Instant::now();
for _ in 0..reps {
  let _ = expr();
}
device.sync().unwrap();

// B: result kept until sync, sync every evaluation.
for _ in 0..reps {
  let t = Instant::now();
  let d = expr();
  device.sync().unwrap();
  times.push(t.elapsed());
  drop(d);
}

// C: all results kept, one sync at the end.
let t = Instant::now();
let outs: Vec<Tensor<2>> = (0..reps).map(|_| expr()).collect();
device.sync().unwrap();

4096x4096 f32, (a * b + c).exp(), 20 evaluations
A drop each result:       346.672µs per eval, 0 launch(es)
B keep + sync each:         1.378ms median,   20 launch(es)
C keep all, one sync:       4.122ms per eval, 1 launch(es)
4096x4096 f32, (a * b + c).exp(), 20 evaluations
A drop each result:       347.037µs per eval, 0 launch(es)
B keep + sync each:         1.306ms median,   20 launch(es)
C keep all, one sync:       4.065ms per eval, 1 launch(es)

        // Hold the output alive across the sync. A lazy backend (fusion) elides
        // a computation whose result is dropped before it drains — so dropping
        // the warmup output would skip the very kernels (and allocations) the
        // capture must reuse, and the capture run would then be the first to
        // allocate them, faulting mid-capture. Keeping `out` until after `sync`
        // forces the closure to actually execute and populate the pool.
        let out = closure();
        let _ = Dispatch::sync(dispatch);
        drop(out);

let (out, d) = device.profile(|| (a.clone() * b.clone() + c.clone()).exp()).unwrap();
show(&format!("[{i}] profile expr"), d);
drop(out);
let (out, d) = device
  .profile_with(ProfileOptions::default().flush(), || (a.clone() * b.clone() + c.clone()).exp())
  .unwrap();
show(&format!("[{i}] profile_with(flush) expr"), d);
drop(out);

[0] profile expr                   no time (device)
[0] profile_with(flush) expr        786.500µs (device)
[0] profile_with(flush) matmul       23.654ms (device)
[1] profile expr                   no time (device)
[1] profile_with(flush) expr        761.417µs (device)
[1] profile_with(flush) matmul       22.694ms (device)
[2] profile expr                   no time (device)
[2] profile_with(flush) expr        758.666µs (device)
[2] profile_with(flush) matmul       22.549ms (device)