Fix n = 0 case for C interface of SSIDS

[mjacobse]

A valid C implementation may return NULL for malloc(0). In that sense, zero-length arrays may be represented by a null pointer in C. When those are passed to the C interface of SSIDS, they are detected as not C_ASSOCIATED, and the corresponding Fortran pointers are nullified. A few of those nullified Fortran pointers are then used illegally without further checks. To fix this, I propose to avoid checking for C_ASSOCIATED altogether by declaring the C arguments as arrays directly.

The problem can be observed for example by setting

int n = 0;
long ptr[1] = {options.array_base};
int *row = NULL;
double *val = NULL;
double *x = NULL; 

in examples/C/ssids.c. A few issues about uninitialized values are reported by valgrind when running the example with these modifications. When additionally setting options.array_base = 0, the example segfaults.

See #82 for a discussion on a different solution that led to this simplified one.

[jfowkes]

Adding an explanation of the non-obvious changes made (see #82 for full discussion):

• Size of crow depends on coptions%array_base in the non-_coord cases
• Could not get rid of fptr and frow altogether, as they may have to be repointed to fptr_alloc and frow_alloc to change C to Fortran indexing
• Had to declare cx1 as assumed-size array, because the size is only known from fakeep%n, which as far as I can tell can only be accessed safely after calling C_F_POINTER on cakeep