Client workload generator

Start making the client workloads more robust.

Overview of this change:

• Cleaned up src/client.cpp
• Multiple workloads can run efficiently from the same workload engine. These are defined in src/clockwork/workload/workload.h and .cpp
• Specific workloads used for specific experiments can be defined in src/clockwork/workload/example.h or in other files
• src/client.cpp lets you specify which workload to run.

Current supported workload types are:

• ClosedLoop: runs concurrency requests in closed-loop
• BurstyClosedLoop: like ClosedLoop, but defined by a burst and idle period. The durations of burst and idle are drawn from poisson distributions, and the BurstyClosedLoop workload is paramaterized by the mean burst and mean idle duration.
• FixedRate: submits requests at a fixed rate. Paramaterized by the rate
• PoissonOpenLoop: like FixedRate, but exact request inter-departure times are drawn from a poisson distribution
• BurstyPoissonOpenLoop: like PoissonOpenLoop, with burst and idle periods similar to BurstyClosedLoop.

Other workloads to add:

• Trace-replay workloads (e.g. from the Azure traces)

include/clockwork/client.h

@@ -29,6 +29,8 @@ public:
 	virtual std::string source() = 0;
 	virtual size_t input_size() = 0;
 	virtual size_t output_size() = 0;
+	virtual int user_id() = 0;
+	virtual void set_user_id(int user_id) = 0;
 
 	/* 
 	Perform an inference with the provided input and return the output.
@@ -45,6 +47,10 @@ public:
 	*/
 	virtual std::future<std::vector<uint8_t>> infer_async(std::vector<uint8_t> &input) = 0;
 
+	virtual void infer(std::vector<uint8_t> &input, 
+		std::function<void(std::vector<uint8_t>&)> onSuccess, 
+		std::function<void(int, std::string&)> onError) = 0;
+
 	/*
 	This is a backdoor API call that's useful for testing.
 	Instructs the server to evict the weights of this model from the GPU.