Background in #1165 and #600.

When the verbose option is enabled, logs are printed on stderr. At the same time, the subprocess might print its output on stdout or stderr. To ensure that verbose logs are always printed in the same order when they interleave with the subprocess output, we are using synchronous I/O (fs.writeFileSync(2, ...)).

Unfortunately, this does not work. That's because, even when stderr is written synchronously, it seems to be buffered by the OS. Since stdout and stderr are using different file descriptors, verbose logs might still be written out-of-order.

At least, that's how I explain it, I could be wrong. The thing that's sure is that, when writing automated tests for this, the out-of-order problem was still clearly present.

The current solution using fs.writeFileSync() has several caveats:

• It uses sync I/O, i.e. is poorer performance-wise.
• As described in fix unicode problems with verbose log #1165, Node.js does not use the same logic with fs.writeFileSync() than with process.stderr.write(). In particular, it does not use some TTY-specific logic, which includes some UTF-8 to UTF-16 conversion on Windows. This leads to weird characters being printed on Windows. In general, I am not surprised fs.writeFileSync(2, ...) leads to weird behavior since this is quite an unusual way to write to a terminal's stderr which Node.js probably hasn't optimized for.
• This is not benefiting from console which has some niceties when printing to a TTY (which verbose logs are likely to) such as more graceful error handling.

With all of this taken into account, this PR is reverting to using console.warn() (as opposed to console.log(), so we still print to stderr) instead of fs.writeFileSync().