Removed unneeded prefetch in IOFiber

[djspiewak]

This was an artifact of our -Xcheckinit days. The net result here is reducing the memory footprint of IOFiber by two full pointer widths, which is actually not a small amount. Surprisingly though, the results are almost exactly a wash for reasons I don't understand... except for parTraverse, which is mysteriously 20% faster now. I reproduced the performance delta across repeated runs and negative tests, so I think it's real, but I do find it fascinatingly bizarre.

Even more interestingly, I would have expected the performance results here to be essentially the inverse of #3471, which saw a statistically-significant degradation from shifting from a megamorphic invokevirtual to a getfield (at the cost of a slightly larger object map). The only explanation I have for why this PR has such minimal impact while that PR was so very noticeable (aside from experimental error) is that perhaps we're right up against an L1 page boundary on my test machine. If that's the case, rebasing #3471 on top of this PR and then re-running should produce a noticeable improvement.

[vasilmkd]

I was planning to do this almost a year ago. I'm glad to see them go. I might have some more ideas for reducing the size of IOFiber even further.

As far as #3471, do be careful about increasing the memory footprint of each IO, it might be a noticeable change for some huge users.