Always refill permits on rate limiter; use governor clock instead of std

[nadav-govari]

Calling available_permits() didn't actually supply an accurate portrayal of capacity, as enough time may have elapsed that a refill is needed. So now a refill is automatically triggered on any call to available_permits() or acquire().

Additionally, the governor crate has a custom implementation of a clock that is faster than std::time. I benchmarked it as follows:

    while benchmark_start.elapsed() < Duration::from_seconds(20) {
        let acquire_start = Instant::now();
        let acquired = rate_limiter.acquire(100_000);
        let latency = acquire_start.elapsed();

        latencies.push(latency);
        iterations += 1;

        if acquired {
            successful_acquires += 1;
        } else {
            failed_acquires += 1;
        }
    }

With the following results:

Metric	code in main	with refill changes, Instant::now	governor::DefaultClock	governor::QuantaUpkeepClock updating every 1ms
Min	25ns	49ns	31ns	25ns
Avg	51ns	51ns	32ns	27ns
P50	51ns	50ns	32ns	27ns
P90	52ns	51ns	33ns	28ns
P99	52ns	52ns	34ns	28ns
P99.9	68ns	72ns	53ns	42ns
Max	45.04µs	147.62µs	94.50µs	26.45µs

Seeing as the benchmark itself used Instant::now, this is probably not scientific, but at these latencies this doesn't really matter.

I went with the default governor clock seeing as an OS thread updating every millisecond is overkill, and the speed of this is not going to contribute to any meaningful bottlenecks enough to optimize this further.