This change addresses two bugs.

1. The existing limit.go implements its own semaphore which ignores the
   passed in context.  This means that the semaphore will block
   indefinitely, even when the context of the request has already
   expired.

2. The existing implementation does not guard against clients which
   invoke RPC methods after the channel underlying the custom semaphore
   has been closed.  This results in panics where the closed channel is
   written to during shutdown, and results in test flakiness.  This
   flakiness is most evident in the op-e2e-http-tests suite.

Along with these fixes comes a test which attempts to demonstrate the
previous bad behavior.  Because these bugs are inherently tied to the
interaction of multiple go routines, the test ends up being a bit
complex, but is well commented and hopefully remains readable.