add a straight forward approach to zero only the bias
this is done by implementing a more general approach to give the
option to zero out any parameter
mod_params and BasicHook now accept a keyword argument 'zero_params',
which will set all parameters with a name equal to or in zero_params
to zero
this is implemented by wrapping the supplied modifier in mod_params
using zennit.core.zero_wrap with a function that will check for the
name, and if matching, not pass the tensor to the real modifier, but
return zeros like the input tensor
supporting rules are Epsilon, Gamma, ZPlus, AlphaBeta, ZBox, WSquare
and Flat
all composites with supporting rules now also have the keyword
argument 'zero_params', which will be applied to all rules
BasicHook now stores the attributes param_keys, require_params and the
new zero_params in a single dictionary attribute param_kwargs; tests
were modified to support this change
BasicHook.copy now keeps the type for copied subclasses; the subclass
init however is never called, therefore setting custom attributes
etc. in the init will require a custom .copy method
updated tests for the appropriate rules and the modified mod_params
updated the example script share/example/feed_forward.py to specify
whether or not the bias should be used
added usage of zero_params to documentation in getting-started and
how-to for rules, composites and canonizers