vecplanes_eval: best implementation avx512
3. what a refine's sums cost
  384 dimensions (a pool of 65536 rows, 24 MB): ns a refine, warm / pool / flushed
                                           portable                 popcnt                   avx2                 avx512
    lionvec_sum8()                265 /   430 /   824      262 /   416 /   822       93 /   387 /   721       92 /   346 /   719
    plane 1 (4-bit query)          37 /   134 /   NaN       10 /   111 /   NaN        9 /   164 /   NaN        7 /   114 /   NaN
    quantized, plane 2             86 /   161 /   250       34 /   135 /   211       23 /   133 /   210       15 /   120 /   196
    quantized, planes 2-3         178 /   269 /   397       65 /   174 /   315       48 /   166 /   299       31 /   137 /   278
    quantized, planes 2-4         279 /   383 /   562       99 /   206 /   432       70 /   190 /   401       47 /   157 /   422
    quantized, planes 2-5         358 /   493 /   645      132 /   255 /   455       93 /   246 /   412       59 /   184 /   424
    quantized, planes 2-6         443 /   587 /   792      168 /   338 /   560      116 /   282 /   515       73 /   190 /   535
    quantized, planes 2-8         627 /   754 /  1097      244 /   404 /   795      163 /   318 /   732      104 /   272 /   724
    bytes, 16-bit query           157 /   364 /   761      163 /   369 /   773       16 /   171 /   573       16 /   151 /   547
    lionvec_dot8()                 73 /   198 /   677       73 /   277 /   683       24 /   149 /   572       22 /   147 /   570
    floats, plane 2               488 /  2207 /  2179      490 /  2221 /  2191      109 /   301 /   276       64 /   200 /   236
    floats, planes 2-4           1490 /  6366 /  6554     1514 /  6338 /  6370      325 /   488 /   644      190 /   322 /   516
    floats, planes 2-8           3669 / 14876 / 14494     3671 / 14514 / 15092      763 /   964 /  1276      444 /   730 /   987
  1536 dimensions (a pool of 16384 rows, 24 MB): ns a refine, warm / pool / flushed
                                           portable                 popcnt                   avx2                 avx512
    lionvec_sum8()               1005 /  1167 /  2107     1009 /  1192 /  1799      368 /   523 /  1500      365 /   511 /  1500
    plane 1 (4-bit query)         148 /   246 /   NaN       34 /   145 /   NaN       26 /   139 /   NaN       19 /   119 /   NaN
    quantized, plane 2            322 /   412 /   602      106 /   198 /   441       56 /   166 /   373       29 /   135 /   357
    quantized, planes 2-3         644 /   739 /  1091      214 /   311 /   789      108 /   237 /   729       58 /   185 /   671
    quantized, planes 2-4         963 /  1138 /  1545      308 /   429 /  1068      166 /   309 /  1001       87 /   223 /   896
    quantized, planes 2-5        1285 /  1478 /  1908      420 /   561 /  1269      221 /   372 /  1159      115 /   254 /  1036
    quantized, planes 2-6        1593 /  1771 /  2262      530 /   749 /  1454      277 /   455 /  1232      143 /   303 /  1109
    quantized, planes 2-8        2251 /  2440 /  2993      726 /   943 /  1674      400 /   575 /  1471      212 /   359 /  1227
    bytes, 16-bit query           618 /   961 /  1769      616 /   943 /  1707       57 /   205 /  1083       57 /   223 /  1064
    lionvec_dot8()                274 /   439 /  1406      271 /   438 /  1384       78 /   241 /  1136       75 /   242 /  1100
    floats, plane 2              2033 /  8750 /  8326     2027 /  8276 /  8284      453 /   545 /   742      251 /   361 /   542
    floats, planes 2-4           6240 / 25639 / 24878     6241 / 24340 / 24534     1373 /  1531 /  1993      756 /   906 /  1404
    floats, planes 2-8          33344 / 58660 / 57739    31005 / 56916 / 57034     3179 /  3377 /  4085     1754 /  2093 /  2738
vecplanes_eval: 0 checks failed
