Include function name and code in the graph debug info.

This patch also makes the other two changes to reduce the overhead of
generating graph debug information:

- reduce the level of stack traces to 3 lines. We found 3 lines is sufficient
  to identify how the node is created at runtime (see the example);

- remove the sorting of the unique file names.

After the change, saving the debug information when training mobilenet v2 can
take as much as 10 mins on a desktop machine. The produced graph debug
information file is about 17.5 M in binary format.

Example:

One of the Conv2D node from the train graph of mobilenet v2:

```
traces {
  key: "MobilenetV2/Conv/Conv2D"
  value {
    file_line_cols {
      file_index: 26
      line: 1156
      func: "convolution2d"
      code: "conv_dims=2)"
    }
    file_line_cols {
      file_index: 4
      line: 182
      func: "func_with_args"
      code: "return func(*args, **current_args)"
    }
    file_line_cols {
      file_index: 28
      line: 278
      func: "mobilenet_base"
      code: "net = opdef.op(net, **params)"
    }
    file_line_cols {
      file_index: 4
      line: 182
      func: "func_with_args"
      code: "return func(*args, **current_args)"
    }
    file_line_cols {
      file_index: 28
      line: 359
      func: "mobilenet"
      code: "net, end_points = mobilenet_base(inputs, scope=scope, **mobilenet_args)"
    }
    file_line_cols {
      file_index: 4
      line: 182
      func: "func_with_args"
      code: "return func(*args, **current_args)"
    }
    file_line_cols {
      file_index: 32
      line: 164
      func: "mobilenet"
      code: "**kwargs)"
    }
    file_line_cols {
      file_index: 4
      line: 182
      func: "func_with_args"
      code: "return func(*args, **current_args)"
    }
    file_line_cols {
      file_index: 45
      line: 93
      func: "network_fn"
      code: "return func(images, num_classes, is_training=is_training, **kwargs)"
    }
    file_line_cols {
      file_index: 5
      line: 165
      func: "build_model"
      code: "inputs, depth_multiplier=FLAGS.depth_multiplier)"
    }
  }
}
```

PiperOrigin-RevId: 233663996