Fix regression in unwinding overallocation on auto cells

Fix regression in unwinding overallocation on auto cells

In https://codereview.chromium.org/1111443003 we started bailing early if about
to overallocate on auto and fixed width columns. The shrink width algorithm wasn't
getting the correct parameters for either of these width types, so our behaviour
when overallocating auto columns regressed.

Fix the parameters passed to the shrink width algorithm so that it works as expected
for auto columns when we have bailed out from allocating width early.

Restore the old behaviour for fixed width columns and work out how to fix it in 
a separate patch.

BUG=498165
Committed: https://src.chromium.org/viewvc/blink?view=rev&revision=197131

[rhogan, 2015-06-12 17:30:48]

robhogan@gmail.com changed reviewers:
+ mstensho@opera.com

[mstensho (USE GERRIT), 2015-06-15 07:25:20]

Need to find a consistent and clean solution, but getting rid of the regression
is pri #1, so lgtm.

[rhogan, 2015-06-15 17:30:22]

The CQ bit was checked by robhogan@gmail.com

[commit-bot: I haz the power, 2015-06-15 17:30:36]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1181643007/1

[commit-bot: I haz the power, 2015-06-15 18:37:53]

Committed patchset #1 (id:1) as
https://src.chromium.org/viewvc/blink?view=rev&revision=197131