This patch attempts to reduce the (stack) memory footprint of
expression classifiers.  Instead of keeping space in each
classifier for all possible error messages that will
(potentially) be reported, if an expression turns out to be
a pattern or a non-pattern, such error messages are placed in
a list shared by the FunctionState and each classifier keeps a
couple of indices in this list.  This requires that classifiers
are used strictly in a stack-based fashion, which is also in line
with my previous patch for revisiting non-pattern rewriting.

R=adamk@chromium.org
BUG=chromium:528697

Review-Url: https://codereview.chromium.org/1708193003
Cr-Commit-Position: refs/heads/master@{#36897}