Pike 8.1.13 (latest git) canonical en-/decoding bug

Hi Pike devs,

While playing around with some code using 8.1.13 from git I ran into 
some really strange behaviour after decoding a mapping that had been 
encoded using encode_value_canonic().
The decoded data turned out to be quite different from what had been 
originally encoded.

I compiled a new Pike from the latest git commit but the issue was still 
occurring, so I set out to narrow down what triggers this and managed to 
produce the following test case.

Some observations:
- The value of m->a->foo and the index/key in m->b must be the same.
- m->a->bar must be above a signed 32-bit int.


int main() {
     mapping m = ([
         "a": ([
             "foo": "potato",
             "bar": 2147483647 + 1, // max of a signed 32-bit int + 1
         ]),
         "b": ([
             "potato": "<-- this should be potato"
         ]),
     ]);

     write("%O\n", m);
     write("================\n");
     write("%O\n", decode_value(encode_value_canonic(m)));
}


Which produces the following output:

([ /* 2 elements */
   "a": ([ /* 2 elements */
       "bar": 2147483648,
       "foo": "potato"
     ]),
   "b": ([ /* 1 element */
       "potato": "<-- this should be potato"
     ])
])
================
([ /* 2 elements */
   "a": ([ /* 2 elements */
       "bar": 2147483648,
       "foo": "potato"
     ]),
   "b": ([ /* 1 element */
       "foo": "<-- this should be potato"
     ])
])


Best regards,
Pontus