Some test results using this PR with PyPy 7.3.0 to compile Iron Horse nml

• about 40% faster for an nml->newgrf compile compared to nml master rev 9fdd4bb
• the Iron Horse makefile has an nfo step, for which the 'Writing nfo' output step is specifically faster by about 85%

Python 3.8 timings are unchanged, as expected.

Much neater than my attempt!

That said, Ijust discovered that PyPy upstream have fixed this in 7.3.1:

Improve performance of io.StringIO(). It should now be faster than CPython in common use cases.

(the changelog categories are confusing, but I've verified that the relevant commits appear in 7.3.1 and not 7.3.0).

That's done by internally using a StringBuffer somewhat like this approach...

Testing on Linux with pypy 7.3.1, there is no measurable difference in performance with this patch; the present code is as fast with 7.3.1 as the patched version in 7.3.0 and earlier.

I think we can rely on 7.3.1, so I'm going to close this - thanks for all the contributions!