Improve `uint_to_hash()`

[flexatone]

This is presently using an simple division by 2, which is not ideal.

Some ideas from Chat GPT:

int64_t hash_uint64_to_int64(uint64_t value) {
    return (value ^ INT64_MIN) & 0xFFFFFFFFFFFFFFFF;
}

int64_t hash_uint64_to_int64(uint64_t value) {
    uint64_t xor_shifted = value ^ (value >> 32);
    return (int64_t)(xor_shifted - INT64_MIN);
}

[flexatone]

This function seems to give the best results:

int64_t hash_uint64_to_int64(uint64_t value) {
    int64_t hashed_value = (int64_t)(value & INT64_MAX);
    if (value & (1ull << 63)) {
        hashed_value = -hashed_value;
    }
    return hashed_value;
}

In Python this is

>>> def h(v):
...     hv = v & np.iinfo(np.int64).max
...     if (v & (1 << 63)):
...         hv = -hv;
...     return hv
... 

We can test that it maps the range correctly with the following:

>>> i64range = range(np.iinfo(np.int64).min, np.iinfo(np.int64).max)
>>> h(np.iinfo(np.uint64).max) in i64range
True
>>> h(np.iinfo(np.uint64).min) in i64range
True

>>> h(np.iinfo(np.uint64).max)
-9223372036854775807
>>> h(np.iinfo(np.uint64).max // 2)
9223372036854775807
>>> h(np.iinfo(np.uint64).max - 4000000)
-9223372036850775807
>>> h(np.iinfo(np.uint64).max // 2 + 1) 
0
>>> h(np.iinfo(np.uint64).max // 2 + 2) 
-1
>>> h(np.iinfo(np.uint64).max // 2 - 2) 
9223372036854775805