Use extended attributes everywhere

[lmolkova]

This is a prototype of open-telemetry/opentelemetry-specification#4485:

• extends standard attributes to include AnyValue
• use them on all public APIs type hints except metrics

TODO:

• Decide if we want MetricAttribute alias
• Tests for resource/instr scope
• More tests for hashing

Benchmarks:

Perf is not a priority - we don't really want people to use complex attributes on metrics, to goal here to show that even if they do, the impact is not huge.

Based on the results, the cost of reporting complex attribute is ~1.5-2 times higher than a standard one.

Name	Min (us)	Max (us)	Mean (us)	StdDev (us)	Median (us)	IQR (us)	Outliers	OPS	Rounds	Iterations
test_histogram_record_10[1]	2.7000	215.5000	3.1023	4.3836	3.0000	0.1000	11;247	322,337.2017	2392	1
test_histogram_record_10[3]	3.0000	222.4000	3.4215	2.5234	3.3000	0.1000	245;3827	292,266.1848	31056	1
test_histogram_record_10_small_mapping_attrs[1]	3.7000	2908.4000	5.3577	30.0136	4.2000	0.5000	32;2907	186,646.7723	16639	1
test_histogram_record_10_array_attrs[1]	3.9000	233.7000	4.5473	3.7219	4.3000	0.2000	400;1218	219,911.5654	32363	1
test_histogram_record_10_json_string_attrs[1]	4.2000	362.5000	4.8095	5.1960	4.5000	0.2000	159;761	207,921.2007	17153	1
test_histogram_record_10_complex_attrs[1]	5.7000	719.1000	6.5912	7.0330	6.3000	0.2000	176;1162	151,717.5248	27625	1
test_histogram_record_10_small_mapping_attrs[3]	5.9000	311.1000	6.7337	4.4705	6.4000	0.3000	191;924	148,505.8706	22676	1
test_histogram_record_10_array_attrs[3]	6.3000	377.6000	7.1968	5.2675	6.8000	0.2000	383;1542	138,950.8709	33004	1
test_histogram_record_10_json_string_attrs[3]	7.0000	181.4000	7.9721	4.2540	7.6000	0.2000	329;1026	125,437.0785	21187	1
test_histogram_record_10_complex_attrs[3]	11.9000	407.1000	13.5603	7.4609	12.9000	0.4000	359;1100	73,744.4408	23530	1

[lmolkova]

TODO: we can use Attributes and let users use complex attributes on metrics without any restrictions. We can also introduce an alias for a limited attribute set.

[adriangb]

I like the idea of allowing it at runtime but blocking it at type checking time. It gives users feedback that they're doing something probably wrong but is easy to override.

[adriangb]

Overall this looks pretty straightforward to me. It goes to show that there's not much of a change for the SDKs to support complex attributes.

[adriangb]

I like the idea of allowing it at runtime but blocking it at type checking time. It gives users feedback that they're doing something probably wrong but is easy to override.

[adriangb]

Maybe keep these tests but remove the assertLogs context manager / i.e. they should pass?

[aabmass]

I like this approach, LGTM