[GraphRuntime] Layerwise timing in debug graph runtime

[hlu1]

The debug_run in the debug graph runtime runs each layer individually and only once, which is not the best way to get the layerwise timing. The PR replaces debug_run with the more accurate run_individual (implemented in #2569) in the python debug graph runtime.

Before:

Node Name  Ops    Time(us)  Time(%)  Start Time       End Time         Shape  Inputs  Outputs  
---------  ---    --------  -------  ----------       --------         -----  ------  -------  
add        myadd  4.54      83.5     21:16:40.659529  21:16:40.659548  (4,)   1       1        
Node Name  Ops    Time(us)  Time(%)  Start Time       End Time         Shape  Inputs  Outputs  
---------  ---    --------  -------  ----------       --------         -----  ------  -------  
add        myadd  1.15      78.68    21:16:40.794619  21:16:40.794770  (4,)   1       1        

After:

Node Name  Ops    Time(us)  Time(%)  Shape  Inputs  Outputs  
---------  ---    --------  -------  -----  ------  -------  
add        myadd  0.071     100.0    (4,)   1       1        
Node Name  Ops    Time(us)  Time(%)  Shape  Inputs  Outputs  
---------  ---    --------  -------  -----  ------  -------  
add        myadd  0.06      100.0    (4,)   1       1        

Note myadd actually only takes about 0.06 - 0.07 us to run, if you warm the cache up first. In the case of a full model, the right way to get layerwise timing is to let the model run continuously with minimal interruptions for a fixed amount of time, and then take the average to get the runtime of individual layers.

@icemelon9 and @antinucleon please review

[tqchen]

Thanks, @hlu1 ! this is now merged