fix(server): strip visual-grounding markup in the tool-suppression branch too

The two streaming call sites read

    if _suppress_markup_display and emit_content:
        ... _suppressed_tool_display_delta(...)
    elif emit_content:
        ... _visual_grounding_display_delta(...)

so a family that suppresses native tool markup took the first branch and
never reached the grounding stripper. zaya_xml is exactly such a family, and
ZAYA-VL emits <|point_start|>tool<|point_end|> as control markup -- which
went out as visible assistant text in both the chat and Responses streams.

The strip now happens inside _suppressed_tool_display_delta as well, applied
before the tool-marker scan so that scan never sees grounding spans it has no
business interpreting. Coordinate-bearing spans keep their inner text as
before; pure control spans are dropped.

tests/test_server.py: 137 -> 139 passing.