i#4197: Add new drwrap post-call scheme: replace retaddr

[derekbruening]

Adds an alternative scheme for achieving a post-call control point
that does not require flushing or shared data structure examination
per-call: replacing the return address with a sentinel.

When the new flag DRWRAP_REPLACE_RETADDR is set, the return address is
replaced with the address of a single return instruction in the client
library, with the real address saved. When a block is seen consisting
of that sentinel instruction, post-call callbacks are called, and then
control is sent to the saved real address using
dr_redirect_native_target().

Adds wrapping tests to drwrap-test.

This new scheme requires restoring return addresses on the stack on
detach or other state translation. Adds functionality to do so, along
with a new test client.drwrap-test-detach.

This requires the client's state restoration event be called for
addresses not in the code cache. Adds such a call.

Adds comments about translation problems with clean call mangling
which is filed as i#4219. The issues seen here are all limited to
traces, so the test works around the problems with -disable_traces.

Tested the core drwrap behavior on ARM and AArch64 but missing general
detach support there (#1578) prevents enabling the detach test there.

Issue: #4219
Fixes #4197

[derekbruening]

Failures are 32-bit build missing a define: will fix now.

[derekbruening]

The Travis Mac job is terminating mid-build. It is strange. It looks like another Travis issue. I filed #4223. There have been transient Travis Mac problems in the past. Hopefully this will go away on its own.