fix(rst): a section's text no longer reads an unwritten byte

[DeusData]

rst_body cuts a reST section's text at 500 bytes. It did so by setting a stop value one past the end and then backing off from out[RST_BODY_MAX] while that byte looked like a UTF-8 continuation byte. That byte was never written, which had two effects:

• Run-to-run variance. Whether the text ended at 499 or 500 bytes depended on whatever the arena memory held.
• Garbage bytes. When a multibyte character didn't fit, bytes that were never written could end up in the text.

Fix: only whole characters are written, and the text ends at the last one written.

Measured

• The same django tree indexed twice with the same main build gave doc_link_candidates rows that differed in about 250 of 11,300. The section text feeds the candidates' TF-IDF scores, so a different cut changes them. With this change, two runs give identical rows and identical text for all 6,594 sections.
• This was found while measuring perf(semantic): exact, cheaper doc-section candidate scoring #2582 on the kernel: the remaining row differences there turned out to come from main itself, not from perf(semantic): exact, cheaper doc-section candidate scoring #2582.

Test. rst_section_text_cut pins the cut at the boundary for three cases: a word that ends exactly at byte 500, a two-byte character that would cross it, and a word cut at a character. The words are spread over short paragraphs on purpose.

• ASan lane: the unwritten byte happens to read as zero in this test process, so the old code passes the test too.
• MSan lane (docker compose -f test-infrastructure/docker-compose.yml run --rm test-msan doc_links_rst, local arm64): with the old rst_body, the test aborts with MemorySanitizer: stack-overflow … nested bug in the same thread. That happened three runs out of three, always at the same pc. With the fix, the suite passes: 9 passed, nothing reported.

Possibly related. The stack overflow appears only when the code reads the unwritten byte, so it looks like MSan failing while it reports that read. The five suites scripts/msan.sh excludes for "stack-overflow under instrumentation" may hide real uninitialized reads the same way. This PR doesn't change those exclusions.