fix: eliminate silent data loss in review pipeline and align docs

[LingMi1]

Summary

Fixes three correctness bugs that silently dropped data or misreported failures, plus aligns stale docs.

Changes

• Remove 6000-byte hunk truncation (internal/prompt/builder.go): hunkTrim silently cut each file's diff at 6000 bytes (~150-200 lines), contradicting the 800-line chunking and dropping the tail of large files. Diffs are now bounded only by ~800-line chunking and the 32 KB prompt cap.
• Check HTTP status in PRDiff (internal/github/client.go): non-200 responses (e.g. 404/500) were read as diff text and misreported as "success, 0 issues". Now returns an error.
• Fail loudly on chunk parse errors (internal/reviewer/reviewer.go): a chunk whose LLM output failed to parse was silently skipped, producing a review with missing coverage. Now the whole review fails instead.
• Docs (README.md, README.zh-CN.md): remove the now-inaccurate "oversized hunks trimmed before construction" claim.

Testing

• go test ./... not run locally (no Go toolchain on this machine). Changes are small and localized; CI / local go test ./... recommended.