Add Korean multicontext Hangul factor dataset with POS tagging (korean_pos_mc)

[kahye]

This PR introduces part-of-speech (POS) tagging to the Korean multicontext Hangul factorization
pipeline, expanding syllable decomposition to 24 lanes.

Key Changes

• New Dataset (data/korean_pos_mc): Adds dataset pipeline, experiment scripts, and lane metadata
for 24-lane Korean multicontext (23 structural Hangul features + 1 POS tagging lane via
kiwipiepy).
• Factorization & Stream Extraction: Updates hangul_factorizer.py, extract_multicontext_streams.
py, and make_multicontext_prompt.py to support POS tag lane generation (--use-pos).
• Testing & Demos: Adds unit tests (test_hangul_factorizer.py) and an end-to-end demo script
(korean_pos_multicontext_demo.sh).
• Training & Visualizations: Minor enhancements to train.py and the 3D vocabulary sphere
visualization app (hf-vocab-sphere) to support the updated multicontext streams.

[kahye]

@gkielian Can you review this PR? Not sure how to add you as a reviewer on this PR.

[gkielian]

Wow this is great! : D

Looking forward to trying this out.

Btw, noticed the input.txt were pretty large files -- would it be possible to instead of placing the input.txt files and other large files -- to instead place generator bash scripts that will create these?

[kahye]

Wow this is great! : D

Looking forward to trying this out.

Btw, noticed the input.txt were pretty large files -- would it be possible to instead of placing the input.txt files and other large files -- to instead place generator bash scripts that will create these?

Great catch. I reverted all the txt changes in this folder since they can be recovered by running get_dataset.sh

[gkielian]

Looks great, thanks! : D