Step1: Optimize direct calls:

Suppose we have a case-lambda as follows

(case-lambda
  [<formals_0> <code_0>]
  [<formals_1> <code_1>]
  ...
  [<formals_n> <code_n>])

First, generate labels for every clause as well as a label for top

(case-lambda <label>
  [<formals_0> <label_0> <code0_>]
  [<formals_1> <label_1> <code1_>]
  ...
  [<formals_n> <label_n> <coden_>])

Now everywhere there is a call to a closure bound variable:
  (funcall x args ...)
We match on the labels of x:
  * If a matching label is found, we transform the call to
    (dircall <label_i> x args ...)
  * If no matching label is found, we transform the call to
    (dircall <label> x args ...)
If we match on a case with ". rest" formals, we emit a call
to "list" on the rest arguments because the label will be
inside the code (after the rest-args are constructed).


Step2: Eliminate passing arg-counts:

All direct calls to inner labels do not require passing an argument
count since it will not be used by the procedure.



Step3:  Eliminate useless case-lambda cases. (not important)

For every label generated, determine whether it is actually used or
not.  

We assume the main label unused until we find:
  * A reference to it in a closure that's not a dircall operator.
  * A reference to it as the target label to a dircall.

We assume an inner label unused until we find:
  * A direct call to it.
  * The main label used.

If a case-lambda has unused clauses, they can be eliminated.

If a case-lambda has an unused main label, the dispatch can be
eliminated.