Add 960ms iq4_xs (lossless, 87MB)

Document q3_k (imatrix) near-lossless 69MB option

Add q3_k (imatrix, near-lossless, 69MB) + imatrix .dat calibration files

Document q8_0/q4_k as recommended formats + quant sweep guidance

Add Q8_0 variants (lossless: token-exact with f16, ~half size, ~1.2x faster on CPU)

Add all 4 streaming chunk variants (f16 + Q4_K) + model card

initial commit