docs: document distribution properties of sample() and sample_in_range()

[jiayaoqijia]

Fixes #4.

Documents the distribution properties of the two sampling methods in DuplexChallenger:

• sample(): Documents that outputs are uniform in [0, P) (the full field), not [0, 2^31). No bias exists for field challenges.
• sample_in_range(): Replaces the vague "Warning: not perfectly uniform" with concrete bias analysis. For the KoalaBear field, masking is perfectly uniform for bits ≤ 24 and has negligible bias (< 1/2) for bits 25-30. The bits = 31 case is excluded by assertion. In WHIR usage, bits is typically ≤ 20 (post-folding domain size), so bias is non-existent in practice.