fix(rag): stop the document cleaner from stripping valid characters ï, ¿, ¾ - #39215

The invalid-symbol filter listed \xEF\xBF\xBE inside a regex character
class. Those bytes are the UTF-8 encoding of the U+FFFE noncharacter, but
the pattern is applied to a decoded str, where they are three separate
code points: U+00EF, U+00BF and U+00BE, i.e. the valid characters ï, ¿
and ¾.

Every ingested document therefore had those characters silently removed,
turning "naïve" into "nave" and dropping the leading inverted question
mark from Spanish questions such as "¿Cómo?". U+FFFE is already handled
by the dedicated substitution on the next line, so removing it from the
character class preserves that behaviour while keeping valid text intact.

The same pattern was duplicated in IndexingRunner.filter_string; both
sites are fixed and covered by tests.