Changing confidence calculation to avoid div by zero

Confidence ranges were coming in too high and had occasional divide-by-zero leading to CHECK failures. Changed the math and the unit tests to satisfy stricter confidence requirements and remove divide by zero.

BUG=95453
TEST=None

Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=99879

[dominich, 2011-09-06 15:38:14]

CHECK was too strict. Clamping instead.

This is a fix for one of the top crashers.

[dominich, 2011-09-06 19:27:43]

Clamping was my gut reaction (patch 1) but then I looked into it further and
realised there was a bug. I took the opportunity to look at histograms I've
collected and found that I wanted to tighten up the constants too.

[sky, 2011-09-06 20:06:32]

LGTM

[commit-bot: I haz the power, 2011-09-06 23:58:29]

The commit queue went berserk retrying too often for a
seemingly flaky test. Builder is mac, revision is 99830, job name
was 7833046-2001 (previous was lost) (previous was lost) (previous was lost)
(previous was lost).