add support for callgraph profiling

[seemk]

New environment variables:

• SPLUNK_SNAPSHOT_PROFILER_ENABLED default: false
• SPLUNK_SNAPSHOT_SELECTION_PROBABILITY default: 0.01
• SPLUNK_SNAPSHOT_SAMPLING_INTERVAL default: 10 (milliseconds)

The profiling is triggered in CallgraphsSpanProcessor which filters the stacktraces based on active traces (i.e. it only keeps stacktraces matching a trace that has been selected for profiling).

Changes to profiling:

• Add the missing profiling.instrumentation.source attribute.
• Use time.monotonic instead of time.time for consistent sleeping.
• Add the ability to pause profiling - which hibernates the profiler thread. Can be resumed again via start. This is used for callgraphs when no traces have been selected for a minute.
• Add support for multiple profiler instances. Context attach/detach wrapping is still done only once, meaning the profiler instances share the thread state mapping.

[pmcollins]

Hi Siim! This is great. Added some comments.

[pmcollins]

Just a nit and a question about shutdown.

[pmcollins]

For clarity, what do you think about this instead:

    is_root_span = not parent_span.is_valid
    return is_root_span or parent_span.is_remote 

[pmcollins]

Should the profiler be shut down here?

[pmcollins]

LGTM. Maybe the shutdown thing is not an issue because it's generally called atexit when the profiler thread will be shutdown anyway. The other comment is just about readability.