Fix ext32 unpacking of the maximum UINT32_MAX-byte payload

[afonsojanu]

Fixes #1086.

The C unpacker tracks how many bytes are left to read for a variable-length value ("trail") by adding 1 to the wire-format length whenever it's an ext (the extra byte is the ext type tag). For ext8 and ext16 that stays comfortably inside a 32-bit range, but ext32's length field can legitimately be the full UINT32_MAX, and tmp + 1 computed in 32-bit arithmetic wraps around to 0 in that one case. The parser then treats the object as a zero-length ext and bails out immediately through template_callback_ext's l == 0 guard, so an ext sitting exactly at the wire format's own maximum size fails to unpack with MSGPACK_UNPACK_PARSE_ERROR, even though it's a perfectly valid encoding.

Fix: widen the internal trail field (and the length parameter the str/bin/ext callbacks receive it through) from unsigned int to size_t, and cast the ext32 length to size_t before adding 1 so the addition itself doesn't wrap on a 64-bit build. This is scoped to the C library specifically (src/unpack.c / unpack_template.h) - unpack_template.h is only ever included from that one file. On a 32-bit build the same boundary case is still unreachable, but that's the same limitation the C++ side already documents as intentional, and a 32-bit process couldn't address a buffer that size in the first place.

Verified with a small standalone reproducer (pack + unpack an ext with exactly UINT32_MAX bytes of data): fails with MSGPACK_UNPACK_PARSE_ERROR on the current code, succeeds with the correct type/size/contents after this change. Also ran a broader manual check across fixext/ext8/ext16/ext32 and str8/16/32 at several boundary sizes (0, 1, 16, 17, 255, 256, 65535, 65536 bytes) under both -Wall -Wextra -Werror and -fsanitize=undefined to make sure the wider type didn't change behavior anywhere else - all green.

Added a regression test (simple_buffer_ext_maxlen in test/msgpack_c.cpp) matching the repro from the issue. It needs roughly 8GB of free memory to allocate the UINT32_MAX-byte source buffer plus the packed copy, so it calls GTEST_SKIP() instead of failing if calloc comes back empty rather than assuming every machine running the suite has that much to spare.