Bypassing capture-output has unexpected behavior when running in REPL

Using the capture/bypass macro works as expected when running tests in the CLI. When running in the REPL (e.g. via cider-jack-in), however, I'm finding that the behavior is not as I would have expected. Namely, the fact that output goes directly to the process stdout/stderr streams means that when running tests in the REPL, output from this macro goes to the NREPL server's output buffer. Ideally this output would instead be written to the writers that NREPL installs in order to capture output, to be returned to the NREPL client and displayed in the Cider buffer.

One potential approach here would be to make the previous bindings for *out* and *err* available via bindings, and have the bypass macro detect and restore these rather than always going to the process file descriptors.

Start by reproducing the capture/bypass behavior through the REPL using cider-jack-in, then compare it with the CLI behavior described in the issue. Trace how the macro handles output and how NREPL installs its writers; done means bypassed output is returned to the NREPL client and displayed in the Cider buffer instead of the server output buffer.