This fixes several (extremely) minor issues in the source code which
were detected as nonconformat by the `-pedantic` flag.

None are serious, and none are likely to resolve any outstanding issues.
The main benefit is that it allows us to apply the `-pedantic` flag into
our testing, which can help to detect future issues.

The issues which have been fixed are described below.

* The `nX` edit descriptor must include the number of forward steps.
  The implicit single step (`1x == x`) is a compiler extension.

* The no-advance line record write token `$` is a compiler extension
  which is not described in the standard.  Non-advancement is handled
  with the `advance='no'` argument.

* Edit descriptors in format statements must be separated by commas,
  even though many compilers will ignore missing commas if there is no
  ambiguity.

* When line continuation tokens are applied to strings, they must appear
  at both the end of the first line and the beginning of the subsequent
  line.  Most compilers do not require this second starting token.

* In function descriptions, if a variable used in the declaration of
  another variable, such as an array index, then it must be declared
  before any other variables refers to it.

  For example, this is invalid:
  ```
     function foo(i0, x)
       real :: x(i0:)
       integer :: i0
  ```
  and `i0` must be declared before `x`.

Statement functions are obsolescent in the current language standard,
and can be redefined as an internal function within a subprogram.

This patch replaces the statement function `psi` (streamfunction) in
`mixedlayer_restrat_general` as an explicit internal function.

The MOM_random generator used bit masks which were set with integer
values.  This is problematic for the sequence 0x8000 0000, because it
must be set with a value of -2**31.

In 4-byte integers, this is strictly not representable in Fortran, which
requires symmetric signed domains for its variables.  Since the upper
bound is 2**31 - 1, the lower bound must be -2**31 + 1, which is larger
than -2**31.  Any value assigned to the 0x80000000 bit sequence is
considered a noncompliant compiler extension.

The current implementation seems to resolve this by using a kind=8 value
(itself problematic, since 8-byte is not assured), but it still requires
assigning this value to a 4-byte integer which cannot (strictly)
represent the value.

This patch averts the whole issue by explicitly setting the bits, and
makes no reference to its integer value.  It leaves the compiler to
decide its interpretation.  And since the variable is only used in bit
operations, there is no ambiguity in behavior.

Note that GCC 9 does not support BOZ conversion from z'80000000' to int,
since it still expects BOZ literals to be within the bounds.  This is
why we use ibset() in place of a literal.  Later GCC versions do not
have this objection.