ddtrace/tracer: priority sampling on by default (DataDog#373)

This change enables priority sampling by default, as well as adds some slight performance improvements to sampling, additional tests and fixes a regression where extracted span weren't sampled.

ddtrace/tracer: added priority sampler (DataDog#371)

* ddtrace/tracer: added priority sampler

* ddtrace/tracer: more tests to ensure context is also updated

ddtrace/tracer: make use of DD_AGENT_HOST and DD_TRACE_AGENT_PORT (Da…

…taDog#362)

This change enables the transport to override host and port information
by reading the environment variables DD_AGENT_HOST and
DD_TRACE_AGENT_PORT.

ddtrace/tracer: add NoDebugStack finishing option (DataDog#355)

Adds a new finish option to the tracer which does not cause the generation of a stack trace on errors for performance reasons in some scenarios.

Additionally, it adds this as opt-in to the grpc integration.

Fixes DataDog#354

contrib/database/sql: only record spans for implemented driver methods (

DataDog#336)

* contrib/database/sql: only record spans for implemented driver methods

* contrib/mongodb/mongo-go-driver: update to latest release

contrib/Shopify/sarama: add support for tracing the sarama kafka pack…

…age (DataDog#296)

contrib/google.golang.org/grpc: fix overwriting outgoing context (Dat…

…aDog#328)

ddtrace/tracer: replace headers when injecting (DataDog#322)

Previously, we were using Add to inject propagation headers. This could
be problematic when a request is forwarded from another service and
already has headers injected. The resulting target may pick up the wrong
context.

gorilla/mux: allow setting custom span options (DataDog#320)

ddtrace/tracer: reset payload after every transport attempt (DataDog#319

)

Previously, we were attempting to save the payload as long as the size
allows it in order to attempt and send it again on subsequent retries to
minimize the loss of traces.

This has proven to be a bad approach because there were cases were the
trace agent did only partial reads from the payload before failing,
leaving the payload in a state where it was impossible to decode using
msgpack (part of the buffer was already read).

This fix is an intermediary easy solution to the problem (we have
experimented with different approaches too, see DataDog#275). We can experiment
later on with retry policies.