Add boot_time cache to improve performance

[dalance]

This PR improves Stat::starttime() performance.
Currently starttime() is slow because it internally calls boot_time() each time.
I think the result of boot_time() can be cached because it is not changed through program execution.

I tried the following benchmark:

#[bench]
fn bench_starttime(b: &mut test::Bencher) {
    b.iter(|| {
        for p in process::all_processes().unwrap() {
            let _ = p.stat.starttime();
        }
    });
}

The result is below:

• Current master

test tests::bench_starttime ... bench: 695,472,752 ns/iter (+/- 35,133,377)

• This PR applied

test tests::bench_starttime ... bench:   7,444,747 ns/iter (+/- 1,975,527)

[eminence]

On my system, the current master branch is about an order of magnitude faster than your system, which is interesting:

test tests::bench_starttime ... bench:  36,540,102 ns/iter (+/- 2,651,593)

And with this patch applied, it's quite a bit faster:

test tests::bench_starttime ... bench:   8,893,658 ns/iter (+/- 917,462)  

Looks like a win!

By my analysis, the call to borrow_mut will never panic because this is a thread-local variable that can never be shared between threads (because RefCell is !Send), so it will never be borrowed simultaneously.