allow direct method setting to support custom layers

[jaybdub]

Currently, converters may only be written for methods with modules visible by torch2trt.py. In addition, converter methods must be specified by full string path, like @tensorrt_converter('torch.relu').

This PR allows the user to specify converters for custom methods.

1. @tensorrt_converter will now automatically determine and import the needed module. For example, @tensorrt_converter("my_package.my_module.my_activation") now works, assuming my_package is installed.

2. @tensorrt_converter will now accept non-builtin methods directly. This helps for locally defined functions, or if it is difficult to determine the full function string name. For example, a local function my_activation can now be converted

   def my_activation(x):
       return 10 * x
   
   @tensorrt_converter(my_activation) 
   def convert_my_activation(ctx):
       # ...