Make ml.lex and ml.grm more closely follow the Definition of SML

The lexer now properly distinguishes between alphanumeric and symbolic
and short and long identifiers.  The parser uses the fine-grained
identifier classes to properly define vid, longvid, tycon, longtycon,
field, strid, longstrid, sigid, and fctid classes.

This fixes a minor bug where MLton erroneously accepted symbolic
identifiers for strid, sigid, and fctid and erroneously accepted
symbolic identifiers before "." in long identifiers.

This also further extends the treatment of `=` as a special identifier
by allowing including `=` in the vid class and equating the con class
with the vid class.  Thus, the program:

    datatype t = = of unit

does not give a parse error, but does give a "Special identifier
cannot be redefined by datata: =." error.

This almost allows `=` everywhere a vid is allowed.  The exception is
that a vidNoEquals class is used (as was previously) in atomic
patterns.  The issue is with

   val pat = exp

and

   fun f apat apat ... apat = exp

phrases, where an `=` in the pattern is difficult to distinguish from
the `=` terminating the pattern(s).

It would be possible to have a set of *NoEqual pattern productions
that excluded the `=` vid, which would be used for these patterns
appearing in declarations.