Add Travis integration, and if possible, use it to check that we can build the example projects.

Support configurable sample sizes and running strategies to obtain samples, such as performing a linear regression on increasing loop counts.

We can then compute a linear regression on the results, as in criterion or core_bench.

The current implementation is a first draft, with the examples acting as integration tests. We need to add unit tests, especially for reporters.

The existing names are confusing. For example, Bencher has a field samples: Vec<Samples>.

Right now, when building with std, we use Instant to measure time, like cargo bench does. Under the hood, on Linux, for example, this uses clock_gettime() with CLOCK_MONOTONIC.

Instead, we should use something like CLOCK_THREAD_CPUTIME_ID, and the nearest equivalents on other platforms. It would be nice for this to be configurable, too, for example, in the case that a benchmark uses threads, and so we might want CLOCK_PROCESS_CPUTIME_ID.

After we have another runner, explore adding a Runner trait and making Bencher generic in it.

For example, support selecting which statistics to report, how to filter samples, serving graphs (possibly via a local HTTP server), JSON output, &c.

Hi Joe!
With both the MIT and ISC licenses in the root directory, but no explanation anywhere, it's unclear what you intend the licensing to be.

The initial implementation is totally reckless wrt panics, casting of numeric types, and floating point arithmetic. Fix this up.

Currently fails with: undefined reference to `__rust_probestack'.

Right now, we have a Reporter struct with a draft approximation of libtest's benchmark output. Remove this, and replace it with a trait. We should also add a line-oriented, delimited reporter that can be used out of the box and fed to Unix utilities.

Also, add #[deny(missing_docs)].