TALVisitor.

  getAttributeNodeNS().

- Move parseAttributeReplacements() and splitParts() out of the class
  (they can be simple functions), so they can be imported by
  TALCompiler and TALInterpreter.

- Remove a redundant key argument from doReplaceLoop() and
  doInsertLoop().

- Add an XXX comment warning that we don't check for z:replace on the
  documentElement.

This really makes a difference!  Some timing statistics (on
test/test2.html, a document with no TAL contents):

talizetree    :  2.190 secs for 10 calls, i.e.  219 msecs per call
printtree     :  2.870 secs for 10 calls, i.e.  287 msecs per call
compiletree   :  0.250 secs for 10 calls, i.e.   25 msecs per call
interpretit   :  0.110 secs for 10 calls, i.e.   11 msecs per call

There are two ways to go from a DOM tree to text (the assumption is
that the DOM tree is already in memory):

talize + print    : total 506 msec
compile+interpret : total  36 msec, or about 14x faster

If we cache the compiled program, we're down to 12 msec, or over 40x
faster.

So compilation + interpretation are well worth it!  (I.o.w., creating
a DOM tree is very slow.)

- Recognize command line options.

  then checking whether it is true.  (Should be slightly faster.)

- Use getAttributeNodeNS() to check for z:omit="" properly.

- The DOMVisitor class is simplified a bit; it doesn't visit
  attributes for you and it doesn't define endVisitElement().

- The CopyingDOMVisitor class is fixed to cope with this; instead of
  endVisitElement() it defines and uses backUp(), with the same
  purpose.  It defines and uses copyAllAttributes() and
  copyAttribute() to deal with attributes.

- The TALVisitor class is refactored so that the processing of METAL
  and TAL attributes is done more orderly (I discovered
  getAttributeNS() :-).

- Used a different way to split parts into semicolon-separated parts
  with doubling used to quote semicolons.

- Sped up the macro indexer by using a recursive function with an
  explicit argument rather than a visitor class.

- Added a slot indexer that works the same way.  (The macro and slot
  indexers have a lot in common, but they're still so small that I
  didn't bother factoring it out.  Later.)

- The runtest.sh script now takes optional command line arguments
  (full pathnames of testfiles) to specify a set of tests to run.

- The timer.py script was adapted to use the new macroIndexer().

returns a tuple (doc, localName) where doc is a DOM tree, or None if
the current DOM tree should be searched, and localName is the name of
the macro inside that tree.

a simple regression test.

I chose semantics that make all variables defined in outer scopes
(including globals) visible as locals, unless overridden by assignment
in inner scope.

http://dev.zope.org/Wikis/DevSite/Projects/ZPT/FrontPage

Added check for 'FORCE_PRODUCT_LOAD' to bits which try to detect whether we're a ZEO client or not in order not to skip product-loading tasks.  This has the consequence that a ZEO client with the "ZEO_CLIENT" env var set will cause product loading tasks to occur iff his "FORCE_PRODUCT_LOAD" environment var is set as well.

a bug that caused ExtensionClasses with __call_method__ hook to fail
to call unbound CMethods.  Thanks!