Allow for pluggable execution_context implementations

[ewhauser]

The current execution_context implementation is hardcoded to use a thread local implementation. Applications running under different threading paradigms, such as asyncio or tornado, require a different context implementation. The only current option available to extend is to monkeypatch the execution_context`.

In order to provide for extensibility, the execution_context could delegate it's method calls to a pluggable implementation, i.e.:

_execution_context_delegate = ThreadLocalExecutionContextDelegate()

def get_opencensus_tracer():
    """Get the opencensus tracer from thread local."""
    return _execution_context_manager.get_opencensus_tracer()

where the ThreadLocalExecutionContextDelegate is just an implementation of the existing execution_context methods. The configuration for the delegate class could either be set directly on the execution_context object (although having a stateful configuration method on the class seems like an anti-pattern) or by introducing a global configuration object like config_integration.

Happy to submit a PR with feedback on the approach.

[ewhauser]

This is related to #234.

[ewhauser]

The current approach I've taken abstracted out the retrieval of the storage of the current tracer into a method and monkey patch that method depending on the host environment:

https://github.com/dronedeploy/opencensus-python/blob/master/opencensus/trace/execution_context.py#L22

[c24t]

#588 should fix this.