[SDK] Tracer::StartSpan benchmark and optimizations

[dbarker]

Fixes # (issue)
The Tracer::StartSpan function currently performs some unnecessary dynamic memory allocation (allocating a SpanContext and potentially a DefaultSpan) and copying of global static shared pointers by constructing temporary span contexts (which call TraceState::GetDefault()).

This latency can be removed without breaking changes. The added benchmark showed a latency reduction of ~11-23% for the range of span creation scenarios (tracer disabled, span with no parent, implicit parent, and explicit parent).

Changes

• Adds a benchmark for Tracer::StartSpan. See baseline vs optimized benchmark results
• Implements several optimizations to reduce dynamic memory allocation and shared_ptr copies (9746077)
  • Adds a Non-recording span implementation with a stack allocated SpanContext
  • Allocate one non recording span per tracer on construction to serve as the fallback noop span
  • Updates the SDK span to take a stack allocated SpanContext
  • Changes the default value of the SpanStartOptions::parent to an empty Context instead of the more expensive SpanContext.
  • use std::make_shared to allocate the span when building with exceptions
  • clean up StartSpan to not instantiate and copy SpanContext if not needed.
  • add GetSpanContext helper to avoid dynamically allocating a DefaultSpan object to get an invalid SpanContext (trace::GetSpan is no longer called).
• Update tracer test cases to check for a non-recording span using the Span::IsRecording method

For significant contributions please make sure you have completed the following items:

• CHANGELOG.md updated for non-trivial changes
• Unit tests have been added
• Changes in public API reviewed

[codecov]

Codecov Report

❌ Patch coverage is 88.75000% with 9 lines in your changes missing coverage. Please review.
✅ Project coverage is 78.15%. Comparing base (c4af827) to head (a815cbd).

Files with missing lines	Patch %	Lines
sdk/src/trace/tracer.cc	87.68%	9 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4248      +/-   ##
==========================================
- Coverage   78.15%   78.15%   -0.00%     
==========================================
  Files         440      440              
  Lines       18741    18762      +21     
==========================================
+ Hits        14645    14661      +16     
- Misses       4096     4101       +5     

Files with missing lines	Coverage Δ
api/include/opentelemetry/trace/default_span.h	36.85% <ø> (+21.06%)	⬆️
sdk/include/opentelemetry/sdk/trace/sampler.h	100.00% <100.00%> (ø)
sdk/include/opentelemetry/sdk/trace/tracer.h	100.00% <ø> (ø)
sdk/src/trace/span.cc	93.58% <100.00%> (ø)
sdk/src/trace/span.h	100.00% <100.00%> (ø)
sdk/src/trace/tracer.cc	81.38% <87.68%> (-6.12%)	⬇️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[dbarker]

The optimizations in this PR result in a 11-23% reduction in latency for StartSpan compared to main.

Benchmark	Baseline (mean ± stddev)	Optimized (mean ± stddev)	Difference %
BM_StartSpanTracerDisabled	6.85 ± 0.025 ns	5.24 ± 0.046 ns	-23.5%
BM_StartSpan	187 ± 1.22 ns	152 ± 2.74 ns	-18.7%
BM_StartSpanWithScope	255 ± 1.09 ns	218 ± 1.75 ns	-14.5%
BM_StartSpanWithImplicitParent	172 ± 0.45 ns	153 ± 0.80 ns	-11.0%
BM_StartSpanWithExplicitParentContext	173 ± 0.34 ns	154 ± 0.90 ns	-11.0%
BM_StartSpanWithExplicitRootContext	187 ± 0.68 ns	148 ± 0.69 ns	-20.9%

[dbarker]

Comparison with main and this PR on span sampler benchmark. The NoopSpanCreation case is where the tracer is enabled but the sampler decides to not sample the span (AlwaysOffSampler).

Benchmark	Baseline (mean ± stddev)	Optimized (mean ± stddev)	Difference %
BM_SpanCreation	235 ± 0.62 ns	205 ± 1.00 ns	-12.8%
BM_NoopSpanCreation	59.3 ± 0.42 ns	27.7 ± 0.14 ns	-53.3%