Add VietOCR recognizer for high-accuracy Vietnamese OCR (lang='vi')

[nhducminh]

Summary

Adds Vietnamese as a first-class OCR language (lang='vi'). PaddleOCR's shared recognizer does not cover Vietnamese tone marks, so pipeline OCR on Vietnamese documents frequently drops/misreads diacritics — e.g. "QUYẾT ĐỊNH" → "QUYT ĐNH". This routes the recognition step to VietOCR (vgg_transformer) while detection stays PaddleOCR, and ships a custom batched decoder that is much faster than VietOCR's stock predict_batch().

Before / after (real Vietnamese PDF page)

	Recognized text
PaddleOCR (before)	Bng 2. Lưng phân hu co, hu co vi sinh cn cung cp theo kt qu phân tích đt
VietOCR (this PR)	Bảng 2. Lượng phân hữu cơ, hữu cơ vi sinh cần cung cấp theo kết quả phân tích đất

Changes

• ocr_language.py, models_config.yml — register "vi" as a public OCR language. It resolves through validate/normalize unchanged and maps its PaddleOCR det/rec/dict to the ch models (detection is used; the rec model still loads as a valid fallback but is bypassed when VietOCR is active).
• pytorch_paddle.py — for lang='vi', load VietOCR and use it for the rec step in ocr()/__call__(); confidence for drop_score filtering comes from VietOCR's own softmax probabilities. Falls back to the PaddleOCR recognizer if VietOCR is unavailable or errors (graceful degradation, no crash).
• vietocr_fast_batch.py (new) — batched VietOCR decoder with a KV-cache (self-attention cache + one-time cross-attention projection over the fixed encoder memory) and per-sequence early-exit. Replaces vietocr's predict_batch()/translate(), which recompute self-attention over the full growing prefix every step and use an unbounded torch.cat() that OOMs on large crop lists. Architecture-guarded: the hand-rolled decode runs only on a post-norm nn.TransformerDecoderLayer stack, otherwise it falls back to an architecture-agnostic decoder — a differently-configured model can't produce silently-wrong output.
• pyproject.toml — optional vietocr extra. Install with mineru[pipeline,vietocr].
• docs/en/reference/vietocr_benchmark.md — rationale, correctness, numbers.

Correctness & performance

• Verified byte-identical to per-image Predictor.predict() — 0 text mismatches across thousands of real crops, for both the KV-cache path and the fallback path.
• Isolated microbenchmark (2000 real Vietnamese line crops, RTX 5080): stock predict_batch() 155 ms/crop → this decoder 9 ms/crop (~17×), 0/2000 mismatches. ~3× on the largest in-pipeline OCR batches.

Notes

• Default OCR language is unchanged (ch); vi is opt-in via -l vi.
• vietocr + PyMuPDF are pulled only by the optional vietocr extra, not by the core install.
• Backward compatible: no behavior change for any existing language.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[nhducminh]

I have read the CLA Document and I hereby sign the CLA

[nhducminh]

recheck