Fix 32-bit overflow in exponent-vector indexing

[dcposch]

On inputs whose total number of terms times the number of variables exceeds INT32_MAX, msolve segfaults while parsing: (pos+j)*nvars is evaluated in 32-bit signed arithmetic, so store_exponent() is called with a negative offset (gdb: pos=-2147477296 at iofiles.c:58, called from get_coefficient_ff_and_term_from_line).

This patch uses 64-bit indexing for exponent-vector offsets and allocations (parser, F4 import, and the variable-order swap).

Minimized reproducer (10000 variables, 214750 terms of x0 over a prime field, ~700KB): unpatched 0.10.1 dies with the backtrace above; patched computes a Groebner basis. make check: 64/64 pass.

[dcposch]

Minimized reproduction example, Python generator:

nvars, nterms = 10000, 214750  # last index × nvars just exceeds INT32_MAX
open("repro.ms","w").write(
    ",".join(f"x{i}" for i in range(nvars)) + "\n105337\n"
    + "+".join(["x0"] * nterms) + "\n")