v3.4.0 — opt-in all-crops line orientation, x1_0 classifier variant, higher det build workspace

v3.4.0 adds opt-in control over text-line orientation, closing #25. CLS_ALL_BOXES=1 runs the 0°/180° line-orientation classifier on every text crop instead of only vertical-looking ones — detection geometry gives each line's axis but cannot spot an upside-down horizontal line, so scans mixing 0/90/180/270 per-line orientations now read correctly (measured cost: about −1% throughput on tiny, −0% on medium). CLS_ONNX / CLS_MODEL additionally accept the shorthand names x0_25 (bundled default) and x1_0, the latter selecting the full-width PP-LCNet_x1_0 textline classifier now shipped in the images (same I/O contract, slightly better flip decisions on hard crops). An explicitly configured classifier that fails to load refuses to boot instead of silently disabling orientation handling.

TRT_DET_WORKSPACE_GB raises the detection engine's TensorRT build-workspace ceiling (default 4 GiB, sized for 16 GB cards) so the medium detector builds at DET_MAX_SIDE_LIMIT=2560 on cards with headroom.

The README gained a Simplified Chinese edition (README_zh.md), a "Getting higher accuracy" section documenting the three accuracy levers with a measured lever-cost chart (every lever combination per tier plus full parse, OmniDocBench-125, RTX 5090), and the v2→v3 migration guide moved to docs/build/upgrading-v3.md.

No breaking changes; all new behavior is opt-in. Verified on a fresh rebuild: unit tests green (GPU + CPU), OmniDocBench-125 accuracy identical to v3.3.0, full-parse 23 pg/s, text-only tiny 271 pg/s in-container.