mca: accumulate statistics of MCA events and export as hw.mca.stats

[mav]

I have doubts such statistics would be usable. Same memory error can be reported in different ways depending how it was found (read, write, scrub, etc). Same time error reported by cache may mean faulty cache or memory error just detected by cache. Excessive simplification like this would tangle things even more. If you wish to have a simple counter, I would only differentiate corrected, uncorrected recoverable and uncorrected unrecoverable (the last should actually end up in panic now, but in some cases could just kill the process). If more detailed information is needed, then some user-space tool like mcelog can get raw data via the sysctl interface.