`Program.Benchmark` should take a `testing.B`

[ti-mo]

Let's go ahead with converting Program.Benchmark() to take a testing.B instead, since it feels pointless to break API twice if we're planning on doing that anyway. When we make this conversion, we can hide Reset() from the caller and report our own metric, e.g. retries to both make sure retries don't go unnoticed and don't need to be handled explicitly by the user.

Then the retries argument also becomes unnecessary as we can take b.N as the value to RunOptions.Repeat.

Originally posted by @ti-mo in #1894 (comment)

[lmb]

@ti-mo taking *testing.B directly has the upside that we can access b.N. There are two downsides:

1. direct dependency on testing, which means the package will be compiled during release builds. Maybe not a big problem.
2. Testing Program.Benchmark from a test is not possible, since we can't construct a testing.B.

I'd probably go with the slightly iffy approach I had originally which introduced type testingB interface. What do you think?

[ti-mo]

testingB is probably the lesser evil, but does that solve point 2?

[lmb]

Yeah, we can make a small fake implementation of it.