This is meant to be used with textlogger for the non-structured log calls to
avoid double encoding.

Textlogger can (but doesn't have to) be used to write klog's unstructured
output from calls like Infof directly to the output stream. This has two
advantages:
- traditional output remains exactly the same as before, which is particularly
  important for multi-line messages because those get quoted otherwise
- performance is better because re-encoding the output string is avoided

In practice, replacing normal klog text output with ktesting lost the time
stamps. This is perhaps not important for unit tests, but for benchmarking
those can be crucial.

Now ktesting uses a stripped down klog header:
- The source code location comes first, because
  that is printed by the testing package.
- Instead if INFO and ERROR, the short I and E are used.
- The useless tid/pid part isn't present.

Example:

    $ go test -v ./ktesting/example/
    === RUN   TestKlogr
        example_test.go:45: I0201 17:58:36.235936] hello world
        example_test.go:46: E0201 17:58:36.236052] failed err="failed: some error"
        example_test.go:47: I0201 17:58:36.236086] verbosity 1
        example_test.go:48: I0201 17:58:36.236110] main/helper: with prefix
        example_test.go:50: I0201 17:58:36.236154] key/value pairs int=1 float=2 pair="(1, 2)" raw={Name:joe NS:kube-system} kobj="kube-system/joe"
        example_test.go:57: I0201 17:58:36.236187] info message level 4
        example_test.go:58: I0201 17:58:36.236209] info message level 5
    --- PASS: TestKlogr (0.00s)
    PASS

Most users won't need this feature. It was enabled by default to keep the API
simple and because the primary goal was unit testing, but benchmarks also need
this and there unnecessary overhead needs to be avoided.

This was removed when introducing sync.Pool in commit
1a1367c because it seemed unnecessary, but an
open issue about the sync.Pool documentation shows that a size limit may be
useful after all.

The intended usage is to replace fmt.Sprintf("%+v") with gomega.format.Object +
YAML support, therefore the only public API change is in the (still
experimental) ktesting.

Internally the additional function pointer gets passed through via a new
Formatter struct. To minimize the impact on klog and textlogger, the
package-level functions still exist and use an empty Formatter.

buffer: restore dropping of too large buffers

ktesting improvements

Being able to change the verbosity at runtime is useful. It is already
supported by the underlying code, ktesting and its Config struct just didn't
expose it.

By embedding *verbosity.VState in Config, users of the package
already had access to V and VModule, but that had two drawbacks:
- not easy to discover
- unclean separate between internal and external API

Providing explicit functions is better.

ktesting + textlogger config api

textlogger write through