nx_quant_q4k.nx

nx_quant_q4k.nx -- hierarchical k-quants (ggml q4_K shape). Improvement over nx_quant_block.nx (q4_0 shape): q4_0: one scale per 32-value block. All 32 values share the same dynamic range; groups with wildly different magnitudes waste precision (small-magnitude groups get coarse steps). q4_K: super-block of 256 values = 8 groups of 32. TWO scales: a SUPER scale per super-block + a LOCAL scale per group. The hierarchy lets each group adapt its precision while still sharing the super-block magnitude estimate -- which empirically halves worst-case error vs q4_0 at near-identical storage. Concrete win for the user's "VRAM optimization realistic tracking": This is the realisation of the Q-002 row -- q4_K + AWQ calibration -- which the doc projected at 1-2% quality loss vs q4_0's 5-7%. Storage per 256 values: 1 super_d (i64) = 8 bytes 8 group_d_local (i64 each) = 64 bytes 128 packed nibbles (256 values) = 128 bytes Total = 200 bytes vs dense 256 * 8 = 2048 bytes -> 10.24x compression Reconstruction: value = nibble * group_d_local * super_d / Q10 Algorithm-led precision: 1. Find super_max = max(|values|) across the 256-value super-block 2. super_d = super_max / 127 (range of effective combined scale) 3. For each 32-value group: a. find group_max = max(|values|) in the group b. group_d_local = (group_max * Q10) / (7 * super_d) (so that nibble of 7 reconstructs to group_max) 4. For each value: nibble = round(value * Q10 / (group_d_local * super_d)) Hierarchy benefit: a group with group_max << super_max gets a SMALL