Add How to use (inference): download snippets, prompt format, greedy protocol

SonaMath-0.5B: rebrand from JunMath + SFT GSM8K full 13.8%

Update eval: GSM8K full test 13.8% (182/1319), greedy, max_new=1024

Fix markdown: replace ~approx with 'approx.' (was rendered as strikethrough)

Clarify SFT-only release vs separate RL research (address overclaim feedback)

Clarify RL theory research as alternative to GRPO-style group optimization

Add architecture-specific theory research section

Emphasize new architecture research for small reasoning LMs

Release JunMath-0.5B research preview (weights + tokenizer + model card)