Atome LM v0.3.0

Initial public release. Routed-ternary tiny language model with bit-exact Python <-> C99 inference engine for microcontroller-class deployment. Apache-2.0. Honest scope: wins at ~60K params (6.31 vs 8.12 ppl, -22%), loses to vanilla FP32 at 944K; MCU parity is QEMU Cortex-M3 only. This release is archived to Zenodo for a citable DOI.