Round instead of truncating in float audio converter

This is mainly to make the tests {float,double}_conversion more robust,
previously they could fail depending on the target CPU type (e.g. with
gcc -m32 -march=i386).
In principle this is now inconsistent with the other converters which
all still truncate, but this doesn't really make a difference in
practice, and making them round would add additional branching and (if
implemented naively) break some other tests, so this is probably fine.

Fixes #274.

libaegisub/audio/provider_convert.cpp

@@ -94,9 +94,9 @@ class FloatConvertAudioProvider final : public AudioProviderWrapper {
 		for (size_t i = 0; i < static_cast<size_t>(count * channels); ++i) {
 			Source expanded;
 			if (src_buf[i] < 0)
-				expanded = static_cast<Target>(-src_buf[i] * std::numeric_limits<Target>::min());
+				expanded = static_cast<Target>(-src_buf[i] * std::numeric_limits<Target>::min() - 0.5);
 			else
-				expanded = static_cast<Target>(src_buf[i] * std::numeric_limits<Target>::max());
+				expanded = static_cast<Target>(src_buf[i] * std::numeric_limits<Target>::max() + 0.5);
 
 			dest[i] = expanded < std::numeric_limits<Target>::min() ? std::numeric_limits<Target>::min() :
 			          expanded > std::numeric_limits<Target>::max() ? std::numeric_limits<Target>::max() :