fix(infer): add repetition_penalty and CLI ngram parameters to prevent text looping

[not-knope]

Description

This PR fixes text repetition/looping issues on dense or structured document pages reported in #55.

Problem:

When processing dense document pages (e.g. forms, government documents), the inference path can enter endless text repetition loops generating 8K–32K redundant tokens until \max_length\ is reached. The root cause is that
o_repeat_ngram_size=35\ only blocks repeating n-grams of exact length 35; shorter repeating phrases (8–15 tokens) fall between 35-gram sliding window boundaries and repeat infinitely.

Solution & Changes Made:

1. **Added --repetition_penalty\ CLI argument and payload support in \infer.py**: Allows passing a repetition penalty factor (e.g. --repetition_penalty 1.1) in \infer_one\ to penalize repeated tokens at the logit level and break short phrase repetition loops.
2. **Added --no_repeat_ngram_size\ and --ngram_window\ CLI flags in \infer.py**: Allows users to configure n-gram size (e.g. \15\–\20) and sliding window sizes instead of keeping them hardcoded at module level.
3. Updated \README.md\ Documentation: Documented --repetition_penalty, --no_repeat_ngram_size, and --ngram_window\ under useful options and added recommendations on mitigating OCR text repetition loops.
4. Added Unit Tests (\ ests/test_infer_repetition.py): Verified default argument parsing, custom CLI flag parsing, and correct payload construction in \infer_one.

Fixes #55