replace bpf_ktime_get_ns with faster bpf_jiffies64

The single call to `bpf_ktime_get_ns` uses about 30% of benchmark time on my machine. It would be a big perf win if we can replace it with something else.

There is a function `bpf_jiffies64` which is probably cheaper, but less accurate. Open questions:

- What is the resolution we can get from jiffies64, and is that enough for our purposes?
- How do we convert jiffies64 to a timestamp?