Improve slow-path performance for unary ops

[robertknight]

Replace iterators with a pattern that uses a fixed number of nested loops. The same approach was previously applied to binary and ternary ops.

Part of #192.

Benchmark:

let mut x = Tensor::arange(0., 1024., None).into_shape([32, 32]);
x.clip_dim(1, 2..30);

let start = std::time::Instant::now();
for _ in 0..1000 {
    x.apply(|x| x + 0.001);
}
let elapsed = start.elapsed().as_secs_f64() * 1000.0;
println!("duration {}", elapsed);

Before: 2.95ms
After: 0.5ms
Without the clip_dim call (making it contiguous): 0.22ms