Add xprof for autotune

[NinaCai]

• For each combination in autotuning, provide the option to generate profiling, and extract the time from profile instead of using wall time.
• Use jit_computation or jitted_computation in the profile, otherwise, use the longest op in the profile.
• Constraint the number of combinations to 10-100. And carefully pick the parameters to sweep.
• Hardcode using 10 iterations + 1 warmup in each script.