Scale test improvements

[graydon]

Two categories of improvement to scale-test:

1. Revive the direct linear regression code as a first-pass (not in log-space) to catch cases that best-fit linearly with no higher-degree exponent term. These cases occur in practice pretty often when someone has a process that's linear "once it gets going" with a sub- or super-linear first point; these presently cause the simplex-algorithm to continue to tweak the polynomial model's exponent to find a curve that fits the extra point, rather than just tilting a line, and it winds up with a bad overall fit (fails, requires more data, etc). Checking a line first (as with the constant early-check) is worthwhile. This also lets us tighten bounds a bit on tests.

2. Complete some work started in 93a832d to make scale-test use the same stats-dir parsing code as process-stats-dir, to use the -stats-output-dir reporting mode of the compiler (not -Xllvm -stats-json) when it's available. This gives us more stats, and also usable scale-testing on non-assert (i.e. end-user) compilers, for the stats we ship in that mode.

[graydon]

@swift-ci please test and merge

[xedin]

I'll try to revert the tests I've modified after changing shrinking, to see if this helps, and will take a look at the code, thanks @graydon !