Use a shared_mutex for read Dictionary operations

[Al2Klimov]

This allows multiple parallel read operations resulting in a overall speedup on systems with many cores.

#9627

real	user	sys
23m41,989s	1056m24,028s	207m46,986s
24m10,306s	1084m32,750s	199m45,335s
24m58,666s	1152m39,816s	168m35,401s
25m16,643s	1156m18,189s	201m20,795s
23m5,677s	1044m19,698s	186m54,275s
24m26,794s	1112m13,336s	184m30,857s
23m5,215s	1049m22,950s	186m35,814s
23m30,688s	1061m56,622s	184m29,615s
23m37,115s	1073m8,501s	185m36,210s
22m32,282s	1037m58,185s	166m31,956s

#9627 + this PR

real	user	sys
21m0,810s	1116m15,330s	3m52,515s
21m32,898s	1150m21,358s	4m11,779s
20m59,009s	1115m39,142s	3m54,868s
21m53,785s	1171m41,773s	3m58,269s
21m18,770s	1128m18,154s	4m1,864s
21m3,465s	1124m55,727s	4m5,976s
21m48,964s	1165m7,774s	4m3,180s
20m10,108s	1077m6,399s	4m9,552s
20m56,285s	1118m42,794s	3m56,123s
20m18,122s	1084m51,305s	4m1,666s

[julianbrost]

Surprising impact on performance, I wouldn't have expected so much contention on dicts. But from the change in system CPU time, this looks quite similar to the corresponding change in namespaces.

[Al2Klimov]

Apply rules eat vars. Vars are dicts.

[julianbrost]

Yes, I just expected less contention as these variables are per object, but apparently there is quite some of it.