Enhenced String report and stacked bar visulization.

[veya2ztn]

1. make the high level profile possible.
   Original profile build the hook for the least leaf module in a model
   For example, only the nn.Linear, nn.Tanh and that elements won't have submodule can get hook.

Old return

MyNet(
  (linear1): Linear(Forward Time: 0.041632ms|Backward Time: 0.014860ms)
  (linear2): Linear(Forward Time: 0.621580ms|Backward Time: 0.009877ms)
  (linear3): ModuleList(
    (0): Linear(Forward Time: 0.054222ms|Backward Time: 0.048355ms)
    (1): Linear(Forward Time: 0.053606ms|Backward Time: 0.050200ms)
  )
  (linear4): Sequential(
    (0): Linear(Forward Time: 0.069707ms|Backward Time: 0.070743ms)
    (1): Tanh(Forward Time: 0.044920ms|Backward Time: 0.039867ms)
    (2): Linear(Forward Time: 0.048994ms|Backward Time: 0.046925ms)
  )
  (linear5): ModuleDict(
    (a): Linear(Forward Time: 0.045158ms|Backward Time: 0.084371ms)
    (b): Linear(Forward Time: 0.046064ms|Backward Time: 0.099857ms)
  )
)

Now it return

It is better to show the whole profile for each level module. Now, the output will become 
MyNet【Forward Time: 2.098945ms|Backward Time: 0.010771ms】(
  (linear1): Linear(Forward Time: 0.041632ms|Backward Time: 0.014860ms)
  (linear2): Linear(Forward Time: 0.621580ms|Backward Time: 0.009877ms)
  (linear3): ModuleList(
    (0): Linear(Forward Time: 0.054222ms|Backward Time: 0.048355ms)
    (1): Linear(Forward Time: 0.053606ms|Backward Time: 0.050200ms)
  )
  (linear4): Sequential【Forward Time: 0.315535ms|Backward Time: 0.229794ms】(
    (0): Linear(Forward Time: 0.069707ms|Backward Time: 0.070743ms)
    (1): Tanh(Forward Time: 0.044920ms|Backward Time: 0.039867ms)
    (2): Linear(Forward Time: 0.048994ms|Backward Time: 0.046925ms)
  )
  (linear5): ModuleDict(
    (a): Linear(Forward Time: 0.045158ms|Backward Time: 0.084371ms)
    (b): Linear(Forward Time: 0.046064ms|Backward Time: 0.099857ms)
  )
)

• The ROOT module MyNet and sub module Sequential get record.
• Notice, ModuleList and ModuleDict still get record. It is designed, since those module may not be used in sequential.
  However, we can also complete it by calling the complete_ignore_cost function in nnprofiler.visulization

2. Add a visulization function that view the profile in stacked bar plot.
   Those timing that won't be recorded will be count as others. See the new README.

[kshitij12345]

Thanks for the update @veya2ztn, will have a look in a day or two.

[kshitij12345]

@veya2ztn this is still on my radar, will have a look soon. Thank you for patience :)