Experiment: Switch to code-driven conversion of CSS unit type names.

Switch to code-driven conversion of CSS unit type names.

The smaller set of CSS unit type names can be resolved without having
to go via a temporary string which is then (hashedly) looked up in a map.
Instead encode the lookup operation in code, avoiding the allocation of
temporary strings and their hashing for each conversion.

Loading content from mainstream sites results in this unit type
conversion operation being performed somewhere in the [5k, 20k] range.

R=
BUG=

[sof, 2015-08-27 05:32:20]

sigbjornf@opera.com changed reviewers:
+ rune@opera.com, timloh@chromium.org

[sof, 2015-08-27 05:32:20]

timloh@: not asking for a review, but if a change of this nature has some
appeal? 

Noticed a larger amount of small strings being created while stylesheets are
being parsed. It seems like allocator churn better avoided.

[Timothy Loh, 2015-08-27 06:05:02]

On 2015/08/27 05:32:20, sof wrote:
> timloh@: not asking for a review, but if a change of this nature has some
> appeal? 
> 
> Noticed a larger amount of small strings being created while stylesheets are
> being parsed. It seems like allocator churn better avoided.

I'd be happy with something along these lines. That said, we need some sort of
unification around this sort of thing, we have too many different ways to parse
strings (make_token_matcher.py, make_element_lookup_trie.py and gperf for
completely generated, as well as custom stuff like this and CSSSelector.cpp
doing a binary search).