NCover report parsing requires a lot of memory

We had a customer with 80MB report, he had to bump up the heap significantly. The problem is that we XPath a parsed document repeatedly, which is both time and memory consuming. We should change the parsing to real single-pass SAX handler based.