OTLP exporter: Fix TypeInitializationException thrown by exporter

[alanwest]

Fixes #1854.

Uses new UnsafeByteOperations.UnsafeWrap method in Google.Protobuf 3.15+ to avoid additional byte[] allocation.

Perf before:

Method	NumberOfBatches	NumberOfSpans	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
OtlpExporter_Batching	1	10000	83.97 ms	1.561 ms	1.533 ms	4166.6667	1500.0000	500.0000	21.29 MB
OtlpExporter_Batching	10	10000	878.07 ms	16.511 ms	27.129 ms	47000.0000	19000.0000	7000.0000	212.88 MB
OtlpExporter_Batching	100	10000	9,006.40 ms	49.816 ms	44.160 ms	509000.0000	201000.0000	89000.0000	2128.82 MB

Perf after:

Method	NumberOfBatches	NumberOfSpans	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
OtlpExporter_Batching	1	10000	83.43 ms	1.487 ms	1.318 ms	4500.0000	1833.3333	666.6667	21.67 MB
OtlpExporter_Batching	10	10000	872.01 ms	13.699 ms	11.440 ms	45000.0000	18000.0000	7000.0000	216.7 MB
OtlpExporter_Batching	100	10000	9,164.78 ms	122.865 ms	108.917 ms	547000.0000	222000.0000	99000.0000	2166.95 MB

Seems there is a bit of an increase in memory allocated... I will investigate why.

[codecov]

Codecov Report

Merging #1873 (113c390) into main (9ebc1ce) will decrease coverage by 0.00%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main    #1873      +/-   ##
==========================================
- Coverage   83.56%   83.56%   -0.01%     
==========================================
  Files         192      192              
  Lines        6163     6150      -13     
==========================================
- Hits         5150     5139      -11     
+ Misses       1013     1011       -2     

Impacted Files	Coverage Δ
...metryProtocol/Implementation/ActivityExtensions.cs	84.93% <100.00%> (-0.85%)	⬇️
...emetry.Api/Internal/OpenTelemetryApiEventSource.cs	82.35% <0.00%> (+5.88%)	⬆️

[alanwest]

@CodeBlanch @cijothomas @reyang

I fairly certain the increase in memory allocated is not due to the code changes here. The new UnsafeByteOperations.UnsafeWrap method avoids an additional allocation of a byte array. The original code was achieving this by reflection which breaks consumers of the Google.Protobuf package version 3.15+.

I suspect that the increase in allocation is related to upgrading our dependency to 3.15.

[reyang]

LGTM.

[CodeBlanch]

LGTM I don't think the change is the reason for the perf hit:

https://github.com/protocolbuffers/protobuf/blob/e9360dfa53f151e64c996823785e0eb6c8118620/csharp/src/Google.Protobuf/UnsafeByteOperations.cs#L78

https://github.com/protocolbuffers/protobuf/blob/e9360dfa53f151e64c996823785e0eb6c8118620/csharp/src/Google.Protobuf/ByteString.cs#L65

That all seems perfectly reasonable to me. Assuming it is flowing through that ReadOnlyMemory<byte> due to implicit cast from byte[].