Fix model card: match actual HF checkpoint (d=512, 8L, 8Q/2KV, ~23M params, ctx=256, FFN=1344)

Fix model card: actual trained model is d=256, 4 layers, 4Q/2KV, ~4M params (was incorrectly listed as 10M)

Fix model card: context_length=256 (not 512), dropout=0.1 (not 0.0) per checkpoint

Add model card with architecture details, provenance, and training metrics

Update model card with architecture details, training config, and usage instructions

initial commit