This adds `zaptest.NewTestLogger(testing.TB) *zap.Logger` and a
level-override variant of it. This will enable us to use non-nop loggers
from tests without spamming the output.

These loggers will not print any output unless the test failed or
`go test -v` was used.

This reverts commit a73cb3f.

This reverts commit be2b5ea.

This declares a TestingT interface which is a subset of the
functionality provided by testing.TB, and changes zaptest to rely on
that.

This also means that zaptest no longer imports the "testing" package.

Also, no explicit string conversion

This tests all log levels for the happy case.

testLogSpy was logging the `args`, not the actual string.

testingWriter should return the length of the original slice, not the
trimmed slice.

This expands the TestingT interface to a much larger subset of
`testing.TB`.

The following were left out:

- `Error` and `Log` were left out in favor of `Errorf` and `Logf`
- `Fatal*` methods were left out in favor of `Errorf` followed by
  `FailNow`
- `Skip*` methods were left out because our test logger shouldn't be
  skipping tests
- `Helper` was left out because not all supported verisons of Go have
  that

Zap's errorOutput is used exclusively to log internal errors. If this
ever gets used in a test, something catastrophic happened and we should
fail the test.

This changes the logger returned by zaptest.NewLogger to implement this
behavior by passing a `WriteSyncer` to `zap.ErrorOutput` that marks the
test as failed upon being used.

To test this, we set up a test logger and replace its core with one that
will always fail to write.

This simplifies the implementation of testLogSpy by removing all
thread-safety from it. It's only used from these specific tests and
never concurrently.

This documents that loggers built with zaptest.NewLogger default to
debug level logging.