Handle seconds=Inf and very high seconds args gracefully (unbounded runtime if samples and evals are specified, otherwise throw)

[LilithHafner]

Fixes #104

• handle seconds=Inf gracefully
• handle very high seconds gracefully
• handle time_ns() being close to the top of its range and overflowing mid-benchmark gracefully (with no errors)
• add tests for all of the above

[LilithHafner]

RegressionTests

• Increased runtime with very low runtime budget
• Decreased correlation between true runtime and reported runtime

Normal tests:

runtime = @elapsed res = @be sleep(.1) seconds=.05
@test runtime < .2 # hopefully this is not going to get too many false positives

fails on 1.6

[LilithHafner]

Lots of little performance regressions and perf test failures, but all are tiny in magnitude so it's worth merging this bugfix. The regression test failures will be marked as accepted once this merges and the non-regression test failures were already flakey.