[experimental] Add SINDI IP U12 + E5M7 storage

[alexanderguzhva]

This is an experimental PR, which tests U12 instead of U16 and E5M7 instead of FP16 for SINDI IP. It trades the QPS & recall for the smaller database size (~25% less for MSMARCO / SPLADE).

This PR is based on top of #1856, so the refinement is also enabled.

The refinement pareto frontier for MSMARCO/SPLADE is the following for the G4 8c32g machine (results are directly comparable with the one from the description of #1856):

sample_id	refine_k	sindi_query_mass	mean_recall	mean_qps
baseline U16 / FP16	1	1	0.99907	1035
baseline U12 / E5M5	1	1	0.99324	828
78	7.81	0.791	0.99324	1401
3	12.37	0.758	0.99317	1412
48	11.39	0.741	0.99309	1511
16	6.78	0.767	0.99299	1544
64	6.34	0.764	0.99297	1570
69	19.75	0.654	0.99287	1656
51	15.73	0.670	0.99281	1695
52	18.85	0.637	0.99274	1756
28	12.26	0.680	0.99272	1761
61	6.16	0.724	0.99269	1785
81	12.84	0.670	0.99268	1792
19	6.53	0.715	0.99262	1815
70	15.67	0.637	0.99259	1857
35	16.36	0.631	0.99254	1858
20	12.63	0.640	0.99235	1956
55	18.52	0.596	0.99195	1957
7	17.46	0.600	0.99191	1976
89	9.97	0.630	0.99158	2139
46	6.92	0.655	0.99109	2146
94	15.65	0.554	0.99017	2307
87	13.11	0.565	0.9898	2375
65	11.31	0.566	0.98898	2477
93	11.66	0.542	0.98779	2625
47	10.88	0.525	0.98596	2806
23	7.96	0.540	0.98352	2941
8	12.42	0.486	0.98332	2965
97	9.12	0.467	0.97467	3457
24	9.67	0.454	0.97295	3503
95	10.38	0.424	0.96663	3688
82	7.29	0.457	0.96552	3798
18	9.21	0.414	0.95977	3932
36	6.79	0.410	0.94413	4398
49	4.51	0.419	0.92675	4778
13	5.03	0.400	0.92249	4887
21	3.65	0.428	0.91597	4905
68	2.42	0.409	0.86192	5650
32	2.24	0.408	0.8484	5764
10	1.39	0.434	0.77887	5827
72	1.10	0.408	0.71428	6466

[sre-ci-robot]

[APPROVALNOTIFIER] This PR is APPROVED

This pull-request has been approved by: alexanderguzhva

The full list of commands accepted by this bot can be found here.

The pull request process is described here

Details

Needs approval from an approver in each of these files:

• OWNERS [alexanderguzhva]

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[mergify]

@alexanderguzhva 🔍 Important: PR Classification Needed!

For efficient project management and a seamless review process, it's essential to classify your PR correctly. Here's how:

1. If you're fixing a bug, label it as kind/bug.
2. For small tweaks (less than 20 lines without altering any functionality), please use kind/improvement.
3. Significant changes that don't modify existing functionalities should be tagged as kind/enhancement.
4. Adjusting APIs or changing functionality? Go with kind/feature.

For any PR outside the kind/improvement category, ensure you link to the associated issue using the format: “issue: #”.

Thanks for your efforts and contribution to the community!.

[alexanderguzhva]

/kind improvement