Fixes panic when byte_offset is in the middle of multi-byte characters
(e.g., inside é). Adjusts to nearest boundary before counting chars.
Verified with French content. All tests pass.

Fixes panics when running `content fix-flaws` on content with multi-byte
UTF-8 characters (e.g., French accented characters like é).

The issue occurred when calculating byte offsets for link replacements:
- `offset - href.len()` could land inside a multi-byte character
- String slicing at invalid boundaries caused panics

Changes:
- Add character boundary validation in `collect_suggestions()` to ensure
  href start offsets are on valid UTF-8 boundaries
- Add defensive checks in `apply_suggestions()` for both start and end
  offsets, skipping suggestions with invalid boundaries instead of panicking
- Add char boundary check in `calc_offset()` as additional safety net
- Add comprehensive tests for multi-byte character handling

This ensures robust handling of international characters throughout the
fix-flaws pipeline (French, German, Japanese, emoji, etc.).

Remove the unnecessary +10 byte margin when calculating search_start
position, which could cause finding wrong instances of duplicate hrefs.
The rfind() search is precise enough without the extra margin.

Additionally, enhance the warning message when an href cannot be located
by including the actual content region that was searched. This makes
debugging much easier by showing what text was examined rather than just
an offset number.

The fix-flaws command was failing to locate and fix broken links when
hrefs contained HTML entities (&#x27; for ', &lt; for <, etc.). This
occurred because:

1. Issue hrefs come from HTML output with encoded entities
2. Suggestion URLs from redirects also contain encoded entities
3. Raw markdown files contain literal characters (', <, etc.)
4. The search/replace logic couldn't match encoded strings against literals

This commit fixes the issue by:

- Adding html-escape dependency to rari-tools
- Decoding both href and suggestion before searching raw markdown
- Decoding href in actual_offset() when calculating positions
- Adding tests for HTML entity handling

Result: fix-flaws now successfully updates 22 French docs that were
previously failing with "Could not locate href" warnings. The fixes
preserve literal characters in markdown (no entities added).

Fixes issues with French SVG documentation and other files containing
accented characters in URLs.