avoid signed negation overflow in integer_times_pow10

[sahvx655-wq]

The int64_t overload of integer_times_pow10 takes the magnitude with -mantissa, which is signed overflow when the mantissa is the most negative int64_t. Reading the overload I noticed the tests sweep numeric_limits<int64_t>::max() but never min(), so I ran that value under the FASTFLOAT_SANITIZE flags and it aborts with parse_number.h:506: runtime error: negation of -9223372036854775808 cannot be represented in type 'int64_t'; in a C++20 constant expression the call does not compile at all. It is not only a sanitiser complaint: GCC 16.1 at -O2 (aarch64) assumes the negated value is non-negative and emits a signed compare (bgt) for the fast-path mantissa bound, so the wrapped value passes the Clinger gate and integer_times_pow10(INT64_MIN, 0) returns 0x1p+63 where from_chars on the same digits gives -0x1p+63.

Taking the magnitude in uint64_t is well defined for every input, and clang emits identical assembly for the overload at -O2, so nothing changes for values that already worked. The negation happens inside the overload, so that is the only place it can be put right: the function is documented to accept any int64_t, and left alone the sign of the result depends on which compiler and optimisation level built the caller. I added min() cases beside the existing max() ones in tests/basictest.cpp; against the old header they fail the value check under GCC -O2 and abort with SIGABRT in the sanitised build, and they pass with the change.