php-check-html-for-indentation calls php-buffer-has-html-tag on every
line indentation, and the function scans the whole buffer from the start.
For a plain PHP file with no HTML tag it always scans to the end, so on a
large file most of the per-line indentation time is spent re-scanning.

Cache the result by buffer-chars-modified-tick and recompute it only
after the buffer text changes. On tcpdf.php (8513 lines, 294 KB) a warm
php-cautious-indent-line drops from about 3.85 ms to 0.26 ms, and the
result still tracks HTML tags added or removed by later edits.

php-check-html-for-indentation returned nil in a polymode buffer, which
made php-cautious-indent-line and -region skip c-indent entirely, so PHP
chunks of poly-php-html-mode were never indented.  In a polymode buffer
php-mode only ever indents its own PHP chunks, so return non-nil and let
indentation proceed.

While here: make php-warned-bad-indent buffer-local and set it
consistently so a warning in one buffer does not silence others, fix the
"seems to contains" prompt grammar, and give php-check-html-for-indentation
a docstring describing its return value.

Add tests covering the two behaviours reworked here: PHP chunks of a
PHP-in-HTML polymode must indent (a regression fix), and a PHP file that
contains HTML must derive to php-html-template-major-mode.

Install polymode and web-mode as development dependencies, and install
dev dependencies in the test target so the tests run in CI.  The poly
test defines its innermode inline rather than depending on the poly-php
package, because poly-php requires php-mode and would install a released
copy from an archive that shadows the php-mode under test.  Both tests
skip when their optional package is unavailable.