Improve performance of user-callbacked sort functions

[nielsdos]

There are 3 improvements that contribute to the performance:

• We don't need to refcount the parameters because zend_call_function already does this when copying to the call frame.
• We can avoid zval destruction logic in the common case that the zval is an IS_LONG.
• We don't need to check for zend_call_function's return value nor if the retval zval became UNDEF, because that would return 0 while php_get_long also does that. We don't need to optimize for the exceptional case.

For this test script:

<?php
function cmp($a, $b)
{
    if ($a == $b) {
        return 0;
    }
    return ($a < $b) ? -1 : 1;
}

$a = range(0,500);
array_push($a, ...range(500,0));

for($i=0;$i<1000;$i++)
	usort($a, "cmp");

On an i7-4790:

Benchmark 1: ./sapi/cli/php comp.php
  Time (mean ± σ):     222.0 ms ±   4.9 ms    [User: 218.5 ms, System: 2.5 ms]
  Range (min … max):   216.1 ms … 234.8 ms    13 runs
 
Benchmark 2: ./sapi/cli/php_old comp.php
  Time (mean ± σ):     264.2 ms ±   7.7 ms    [User: 258.6 ms, System: 1.9 ms]
  Range (min … max):   256.9 ms … 284.2 ms    10 runs
 
Summary
  ./sapi/cli/php comp.php ran
    1.19 ± 0.04 times faster than ./sapi/cli/php_old comp.php

On an i7-1185G7:

Benchmark 1: ./sapi/cli/php comp.php
  Time (mean ± σ):     147.0 ms ±   2.5 ms    [User: 144.2 ms, System: 2.8 ms]
  Range (min … max):   144.7 ms … 154.6 ms    19 runs
 
Benchmark 2: ./sapi/cli/php_old comp.php
  Time (mean ± σ):     184.3 ms ±   1.6 ms    [User: 180.7 ms, System: 3.3 ms]
  Range (min … max):   182.7 ms … 188.9 ms    16 runs
 
Summary
  ./sapi/cli/php comp.php ran
    1.25 ± 0.02 times faster than ./sapi/cli/php_old comp.php

Note that not only usort improves but it's the easiest to show.

[iluuu1994]

I know this is still wip, but the current code looks correct so far.

[iluuu1994]

Has nothing to do with your code, but: What a weird macro name. 😅

[nielsdos]

I'm going to leave it at this for now. Improving further is possible by handling the deprecation logic slightly smarter but that only gives a few milliseconds improvement and does not seem worth it given that the code will be gone in the near future anyway.
So no changes since the review except for me adding a description to the PR with benchmarks and an explanation.

[bukka]

LGTM