Deep refactoring of named expressions about batchflow

[roman-kh]

Remove F and L expression

Since pipeline has several options to call an external function, we might get rid of F and L expressions, thus simplifying the API and free letters for new expressions.

pipeline.call(external_func, B(), V.some_var, dst=V('other_var'))

pipeline
.add_namespace(some_module)
.external_func(B(), V.some_var, dst=V('other_var'))

So instead of
pipeline.train_model('model', feed_dict=F(make_data)(B.images, V.some_var) (1)

you will write:
pipeline
.call(make_data, B.images, V.some_var, dst=V('feed_dict'))
.train_model('model', feed_dict=V.feed_dict)

or even
pipeline
.make_data(B.images, V.some_var, dst=V('feed_dict'))
.train_model('model', feed_dict=V.feed_dict)

While the first version (1) is shorter, the proposed change might seem less obscure and implicit.

[roman-kh]

Separate a computed NE and a reference NE
While this works well for C, D, B and V, the logic breaks when it comes to R, P, F.
For instance, when you write R('normal', 0, 1), you rather want to compute the values and pass them, not the NE itself.
As a result, when you see NE(...) you still don't know whether the values should be passed or NE.
Consequently, the action might need to recompute all the parameters, which is inefficient.
So, let's use W as a container to pass NE further. And later we might also add parameters for @action decorator to specify arguments which should be passed "as is" and not be computed in the pipeline.

Remove F and L expression
Sometimes L might be helpful to pass a callable with params, so it is a declarative version of partial.
So I'll remove F and rename L to F for clarity. :-)