On a buffer that is one very long soft-wrapped line (e.g. a minified JS
bundle), PageDown from the top jumped the caret straight to EOF, and the
following Up jumped it back to the document start.

handle_page_motion judged "did the viewport scroll?" by comparing only
top_byte. For a single logical line, scrolling advances
top_view_line_offset while top_byte stays 0, so the check always saw "no
movement" and fell through to the logical-line MovePageDown handler, which
clamps the caret to EOF. The off-screen caret then defeated the wrap-aware
MoveUp intercept, so Up fell through to byte-based MoveUp and landed at
byte 0.

Detect viewport movement via the (top_byte, top_view_line_offset) pair, and
land the caret at the source byte of the visual row now shown at the top of
the viewport (new Viewport::top_visual_row_source_byte, which maps the
wrap-segment offset back to a buffer byte). The helper defers to top_byte
when plugin soft-breaks / virtual rows touch the line, leaving plugin and
no-wrap behaviour unchanged.

Reproduced by tests/semantic/page_motion_single_long_line.rs against a new
single-long-line fixture: without the fix PageDown lands the caret at EOF
and the following Up at byte 0; with the fix the caret stays roughly one
page down.