Lots of micro-performance adjustments using conventional Python

techniques of reducing the number of lookups and avoiding method
calls.

More interestingly, the dispatch machinery has changed, especially
TALInterpreter.iterpret().  Each instruction is now composed of a
2-tuple of the form (opcode, (arg1, arg2, ...)).  This is important as
it allows more efficient unpacking and construction of the argument
list for the call to the handlers for individual bytecodes.

The bytecode handlers are also located differently.  Instead of using
the opcode string to construct a method name, retrieve a bound method,
and then calling it via apply(), a dictionary of handler functions is
created as part of the class definition.  Handlers are retrieved from
this dictionary by opcode string (avoiding string concatenation and
bound method creation in the inner dispatch loop).  The handlers (now
functions rather than methods) are then called using apply(), creating
the arguments to the handler by tuple concatenation, which is faster
than creating the bound method and then using the slower
method-calling machinery.  Temporary variables are avoided whenever
possible.

The test for "debug mode" in TALInterpreter.interpret() was moved out
of the dispatch loop; it is used to select one of the two versions of
the loop.

Support has been added for two new bytecodes:
  rawtextColumn -- used when the TALGenerator can determine the
                   resulting column number ahead of time (the text
                   includes a newline)

  rawtextOffset -- used when the TALGenerator cannot determine the
                   final column of the text (there is no newline).

These new bytecodes allow the interpreter to avoid having to search
for a newline for all rawtext instructions -- the compiler is able to
determine which of these two cases is appropriate for each rawtext
chunk -- the old 'rawtext' instruction is no longer generated.

Re-phrased some conditions in if statements to allow more
short-circuiting.