Look up nth-child rules for table rows and cells directly instead of running a regex over every stylesheet key #526

Part of #520 (step 6). Changes nothing on the page.

For every TR, TD and TH, CssMerger runs preg_match('/TAG>>SELECTORNTHCHILD>>(.*)/', $key) to find the nth-child rules. It does so over every key of CssManager::$CSS, over the top level of the cascade tree and over the current table cascade level (src/Css/CssMerger.php:530, :742, :850). That is three passes of regexes over the whole stylesheet for every cell.

Measured on gravitypdf at b6a07921, PHP 8.5.4. A table of 400 rows and 5 columns, and a stylesheet of class and descendant rules, none of which match a cell. WriteHTML() time, mean of 3 runs with a hot cache:

Measure again cold and hot, averaged, before quoting these.

In the same code, CssParser::processCssSelector() counts maxClassDepth over every class in a selector, not per compound (src/Css/CssParser.php:242-251). .a .b .c p sets it to 3 although no compound has more than one class, and rules that can never match (.x:not(.y):not(.z)) count too. Every element with classes then builds combinations up to that size.

Fix.

• At parse time, record the nth-child formulas used for each of TR, TD and TH. Skip the lookup when there are none; otherwise look up each TAG>>SELECTORNTHCHILD>>formula key directly in each node.
• Count maxClassDepth per compound, and only for rules that are stored.

Tests. nth-child rules still apply as before, from the stylesheet, from a descendant rule and from a table cascade level. The snapshot suite passes with no fixture updated.

Related: #522 changes how nth-child keys are formed, and #528 changes the index they are matched against.