nx_q5_k_to_f32.nx

nx_q5_k_to_f32.nx -- bits-up Q5_K dequantization producing IEEE 754 binary32 outputs. Q5_K = the (ggml-correct) Q4_K structure + a 5th bit per value from qh. Real Q4_K_M models store attn_v / attn_output / ffn_down as Q5_K, so this is required to load them sovereignly. ggml block_q5_K (176 bytes / 256 values): [0..2) d f16 super-scale [2..4) dmin f16 super-min [4..16) scales 12 bytes = 8x(6-bit sc) + 8x(6-bit m) (get_scale_min_k4, same as Q4_K) [16..48) qh 32 bytes = 1 high bit per value [48..176) qs 128 bytes = 4 low bits per value ggml dequantize_row_q5_K layout: 4 GROUPS of 64; group g feeds sub-block 2g (low nibble, 32 outputs) and 2g+1 (high nibble, 32 outputs); the 5th bit is qh[l] tested with u1=1<<(2g) (low) / u2=1<<(2g+1) (high). value = d*sc*(q4 + (bit?16:0)) - dmin*m (q5 in 0..31, min subtracted -- same form as Q4_K). license_tier: ORIGINAL