BUG: raise helpful errors for unsupported pyarrow read_csv usage

[jbrockmendel]

Follow-up to the pyarrow-engine xfail cleanup (GH-65859): the remaining read_csv pyarrow xfails were cases where unsupported usage produced a cryptic exception or silently-wrong results. This PR makes them raise informative errors (or behave like the other engines where the fix was simple) and converts the corresponding xfails to assert those errors.

• list-valued header (MultiIndex columns, unsupported by pyarrow.csv) now raises ValueError: header argument must be an integer or None when using engine='pyarrow' instead of TypeError: an integer is required; singleton header=[N] now behaves like header=N
• names together with an integer header was silently ignored; the header row is now discarded and names applied, matching other engines (including implicit-index handling when names is shorter than the file width)
• too many names (or a names/usecols length mismatch) now raises "Number of passed names did not match number of header fields in the file" like the python engine
• na_filter=False was silently ignored (NA filtering happened anyway); it now raises the standard unsupported-option ValueError
• tuple names now produce MultiIndex columns like other engines

~30 xfailed tests converted to the raise-assert pattern; 11 tests now pass outright and lost their xfails.

Note: touches the same areas of arrow_parser_wrapper.py as GH-65859, so whichever merges second will need minor conflict resolution.

🤖 Generated with Claude Code