reimplement whitespace handling for standalone tokens

Whitespace handling was previously done by a post-processor on the
token stream produced by the lexer. We reimplement it as
a post-processor on the *output* stream after rendering. This means
that information from the fully-parsed AST can be used to make
rendering decisions.

For example, non-rendered sections (because their corresponding json
data is empty) are now handled as standalone tokens. This gives
nicer output, consider for example:

    {{#foo}}Foo is present.{{/foo}}
    {{^foo}}Foo is absent. {{/foo}}

This used to produce two lines of output, one being empty. Now this
only produces one line of input (if 'foo' is a scalar or object, not
a list).

(Interestingly, such examples are not tested in the Mustache
specification testsuite. Both the previous implementation and the
current implementation pass all the specification tests, which
suggests that the tests are not comprehensive enough.)

Note: The real motivation for this change is not inverted sections, but
rather indentation and whitespace handling for partial parameters --
when implementing template inheritance.

bin/test/errors/parsing-errors.t

@@ -111,7 +111,7 @@ Mismatch between section-start and section-end:
   $ echo "{{#bar}} {{#foo}} {{.}} {{/bar}} {{/foo}}" > $PROBLEM
   $ mustache foo.json $PROBLEM
   Template parse error:
-  File "wrong-nesting.mustache", lines 1-2, characters 41-0:
+  File "wrong-nesting.mustache", line 1, characters 32-33:
   Section mismatch: {{#foo}} is closed by {{/bar}}.
   [3]
 

bin/test/errors/render-errors.t/run.t

@@ -13,7 +13,6 @@ reference.json and reference.mustache work well together, there is no error.
 
   The variable "name" has value "Some Name".
 
-
   Last line.