Especially when dealing with big files the performance is quite bad. This
change speeds it up a little by using an unordered map to find the elements
for certain offsets instead of using linear lookup.
According to callgrind the modified functions where one with the biggest
own cost. The function updateRelativeOffsets() was actually the 2nd costly
function just below __memcpy_avx_unaligned_erms.
Now TagParser::EbmlElement::internalParse(TagParser::Diagnostics&) is quite
high as well as standard IO stream functions.