Make the parser use file memory-mapping on Windows too.

[Scylardor]

Some thoughts on this:

• Maybe I should have changed the PBRT_HAVE_MMAP define name. But since mmap doesn't exist on Windows, still makes sense.
• I kept the "slurping all characters from file" method if we don't have mmap AND we aren't on Windows (never, basically ?). But otherwise, if we failed to memory-map the file, we error and exit. Perhaps it would be best to fallback on the manual file open and read in all cases if memory mapping fails.

I could only test on my personal Windows 10, sorry!. But as it's pretty classic windows API, I expect this to work smoothly with older versions of Windows.

[Scylardor]

I wrote a fix for the tests because the parser wasn't checking for '\r' in comments (\r\n line endings on windows) so it had a \r after what it expected.

And the other issue was that the parser test failed to remove the test file because the tokenizer was still holding to its mapping view of the file. So, I enclosed it in a scope to destroy it before attempting to remove the file.

[mmp]

This looks great--thanks for doing this!

I'll hold off merging until all of the appveyor builds finish successfully; I assume the current run will be good.