docs: fix GitHub suggestion-count configuration examples

[teachershuang]

Fixes #3331.

Use num_code_suggestions_per_chunk in the three GitHub installation examples. The previous keys left the actual setting at its default. Values and configuration mechanisms are unchanged; no runtime, default, dependency or workflow changes.

Validation (Windows, Python 3.12.13, frozen dependencies, uv 0.12.10):

• Loaded the two environment examples and parsed TOML override with the repository's Dynaconf settings: old examples read 3/3/3; corrected examples read 6/8/6.
• Configuration-reference and documentation-navigation tests: 6 passed.
• Touched-file pre-commit hooks passed; independent Codex diff review found no issues.
• MkDocs build succeeded; checked all three keys in the generated GitHub page. Existing unrelated site/link warnings remain. Terminal capture:

Environment PR_CODE_SUGGESTIONS.NUM_CODE_SUGGESTIONS_PER_CHUNK: expected 6, actual 6
Environment pr_code_suggestions.num_code_suggestions_per_chunk: expected 8, actual 8
TOML num_code_suggestions_per_chunk: expected 6, actual 6
Documentation built in 79.52 seconds

Broader Windows unit run: 7697 passed, 42 failed, 5 errors, 33 skipped, 1 xfailed, 1 deselected (142.79s). Cleared proxy variables only in the test process to avoid the host's SOCKS/missing-socksio issue. All 47 failed/error test IDs also fail on the unmodified parent (exact failed-test rerun: 42 failed / 5 errors); categories include CRLF/GBK, unsupported Windows filesystem/POSIX operations and cleanup timing assertions. No unrelated runtime fixes included.

The deselected case is test_skills_loader.py::TestPathExpansion::test_tilde_in_path_is_expanded: it patches only HOME, while Windows expanduser uses USERPROFILE, causing discovery to traverse the actual user directory. Stopped the original run and isolated reproduction at that case, then explicitly deselected it for the broader run. It is not counted as passing. Restored the exact committed docs and reran the focused checks afterward.

No live provider/e2e run; no executable lines changed, so diff coverage is not applicable. Upstream Build-and-test, pre-commit and CodeQL all await workflow approval with zero jobs. Keeping Draft; local Windows results are not Linux-container CI validation.

Investigation, editing, validation and independent review are AI-assisted.