`strlen(buf)` is not guaranteed to return the size of the buffer, as it
might contain some null bytes in the middle.

Fixes htacg#1001

The loop was introduced 91f29ea when switching to a non-recursive
algorithm. `InlineDup` should not be called when `ParsePre` restarts.

This should fix the oss-fuzz build failure @ https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=36721

NormalizeSpaces decodes and re-encodes UTF-8 characters while looking to
replace non-breaking spaces with regular spaces. When the UTF-8 decoding
hits an error, a replacement character (0xFFFD) is returned and
re-encoded as a 3-byte UTF-8 character. In some cases, this increases
the size of strings, leading to writing past the end of the allocated
buffer.

Fixes https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=13191.

The mmaped IO implementation on windows does not always increment `pos`
on each `getByte`. It relies on an incrementing pointer `iter`, and
update `pos` only when a new chunk gets maped.

However, mmaped_eof was only considering `pos` and not `iter`, and
therefore was sometimes incorrectly returning false until a later call
to `getByte` would update `pos`.