have two main actions: with unroll and without, for no unroll icrease the step by 1 in pilot (not *2)

[adamsitnik]

I am currently working on porting CoreCLR to BenchmarkDotNet.
CoreCLR contains a lot of "benchmark game" benchmarks. These are long-running benchmarks.

After #760 the time it takes to run the benchmarks has greatly improved (ML.NET benchmarks from 20 to 2 minutes). However, I have found two extra scenarios for improvement:

1. The time to execute benchmark is slighly less than IterationTime. Let's say it's 0.4s when the IterationTime is 0.5s. In that case we invoke it 16 times per iteration (default unroll factor value). It's too much, we should invoke it just once.
2. The time to execute benchmark is just few times less than IterationTime. Let's say it's 0.1s when the IterationTime is 0.5s. In that case we invoke it 16 times per iteration (default unroll factor value). It's too much, we should invoke it few times to fit within IterationTime .

My changes are quite simple:

1. If the time to execute benchmark is slighly less than IterationTime, run it once per iteration. No Pilot needed.
2. If the time to execute benchmark is just few times less than IterationTime, start the Pilot with MinInvokeCount = 2 (we know 1 is not enough), and increment it with every pilot iteration (not double, it's typically too much for these benchmarks).

Example:

[Benchmark]
public void Sleep100() => Thread.Sleep(100);

IdleJitting  1: 1 op, 803135.91 ns, 803.1359 us/op
MainJitting  1: 1 op, 100786001.20 ns, 100.7860 ms/op

Pilot        1: 2 op, 200567013.38 ns, 100.2835 ms/op
Pilot        2: 3 op, 301861923.92 ns, 100.6206 ms/op
Pilot        3: 4 op, 401721624.60 ns, 100.4304 ms/op
Pilot        4: 5 op, 503124019.07 ns, 100.6248 ms/op

MainWarmup   1: 5 op, 503016250.03 ns, 100.6033 ms/op

[AndreyAkinshin]

LGTM