feat: group output by file to reduce noise

[mit-d]

Summary

• Findings are now grouped per file with a compact line range header (e.g. file.txt:1,4-80,90:)
• Context lines shown once with multi-caret markers (^ normal, ! dangerous, ? confusable)
• Identical context lines deduplicated (e.g. repeated box-drawing rows shown once)
• Codepoint listing collapsed with (xN) counts, same codepoint found by both check_file and check_confusables merged to the more informative label

Before (100-char box-drawing line): 300+ lines of output

After: 6 lines

file.txt:1:
  └──────────────────────────────────────────────────────────────────────────────────────────────────┘
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  U+2500 BOX DRAWINGS LIGHT HORIZONTAL [So] (x98)
  U+2514 BOX DRAWINGS LIGHT UP AND RIGHT [So]
  U+2518 BOX DRAWINGS LIGHT UP AND LEFT [So]

Found 100 non-ASCII characters in 1 file

Test plan

• All 235 existing tests pass
• New tests for _compact_ranges, _build_caret_line, _format_codepoint_entry
• Manual verification with box-drawing, smart quotes, dangerous, and confusable characters
• Quiet mode, no-color, and file-read-failure edge cases covered