I don't have a recommendation on a good memory tester. There are a few old ones that run in DOS but I don't think they are particularly effective. When things are really bad I break out the DRAM testing device which can test 64K and 256K DRAMs; it works pretty well, but it doesn't work for my later systems with use SIMMs. If the system is new enough memtest86 is a good alternative.
I doubt that your IRQ configuration has an impact on the checksum errors, other than the general observation that RealTek engineers were not taking much pride in their work. ;-0
I uploaded the updated FTP.EXE file to http://www.brutman.com/mTCP/ftp.exe . It just makes it easier to see the number of checksum errors by printing out a line of statistics at the end of a program. I've tested it as best I can, but since I'm not getting any checksum errors it is always 0.
The paper you linked was a good read. Not detecting a checksum error every 1 in 16 million packets can seem like a lot of data, but even my short sessions transferring I was receiving on the order of 100,000 packets and generating an equal amount. For a real machine like an ISP gateway 16 million packets goes by pretty quick. The paper is 20 years old now so the scale of the problem might be different; hardware might be more reliable but we send many more packets now.
Bookmarks