drmemtrace address elision ignoring non-memref writes, leading to incorrect addresses

PR #3977 for #2001 put in elision of recorded addresses we can deduce in post-processing. But it turns out that non-memory-accessing instructions are not properly considered when looking at register mods. The tool.drcacheoff.burst_traceopts test which compares an optimized to un-optimized trace is just missing this pattern of modification.

I made a tool to measure how often this happens and the ratio on most apps is about 1 in 500 data addresses being wrong (typically they'll be off by some small constant depending on the reg mod in between).