fix(compiler): support first-class callable of internal functions

[Tinywan]

Scope: first-class callable handling for internal functions only.

This change skips internal-function argument-count validation when PHP 8.1 first-class callable syntax is used, because foo(...) creates a Closure rather than invoking foo.

It also skips mathematical-function return-type optimization for that syntax, preventing access to VariadicPlaceholder::$value, which is absent for the placeholder argument.

Validation: callable-only commit replayed on current master; git range-diff and git diff --check pass; php vendor/bin/phpunit phpunit/src/AssignTest.php passes (34 tests). Regression coverage is in builtin-multi-arg-callable.phpt and math-function-callable.phpt.

[matyhtf]

The first-class callable fix is valid and can be kept. The two new callable PHPT tests pass, and the existing AssignTest suite remains green when only that commit is applied.

The inferred-object reassignment change cannot be merged in its current form:

1. reassignInferredObjectVar() is not limited to mutually exclusive branches. It permits any inferred local to change between unrelated classes, including ordinary sequential assignments.
2. It breaks the existing negative contract in AssignTest::testAssignClass. This code is currently required to fail, but the PR makes it compile:

$obj1 = new stdClass();
$obj2 = new ArrayObject();
$obj1 = $obj2;

3. Clearing objects / stableObjects while walking the second assignment is source-order mutation, not a control-flow type merge. It implicitly weakens an established object type to generic object, contrary to TypePHP's fixed inferred-type design.
4. The new PHPT uses conditions whose results are already statically known (is_string($a) where $a is a string, and the same check where $b is an array), so it does not demonstrate a sound runtime branch merge.

Please remove the second commit from this PR and keep this PR focused on first-class callable handling. Code that intentionally stores unrelated object classes should opt into dynamic storage with std::any().

If branch-aware object type merging is added later, it needs to be implemented in the SSA/control-flow merge stage, with dedicated tests for both runtime branches, rather than by widening the variable during assignment code generation.

[Tinywan]

The inferred-object reassignment change has been removed from this PR and preserved on the WIP branch: wip/inferred-object-reassignment-pre-ssa. I will only revisit that behavior with a redesign at the SSA/control-flow merge stage.

[matyhtf]

The first-class callable check needs to run before all function return-type optimizations. Currently detectTypeOfExpr() checks STREAM_FUNCTIONS before the generic callable-to-object branch, so stream-returning builtins such as fopen(...) are still inferred as Type::STREAM instead of Type::OBJECT.

A valid example fails to compile:

function useOpener(Closure $open): void {}
useOpener(fopen(...));

The compiler reports: Cannot re-assign variable from php::Stream to php::Object.

Please move an isFirstClassCallable() / Type::OBJECT short-circuit ahead of both the math and stream return-type handling, and add a regression test using a stream builtin (for example, passing fopen(...) to a Closure-typed parameter).

Validation performed on the exact PR head: the compiler built successfully; the 34-test PHPUnit slice, both new PHPTs, and all existing placeholder PHPTs passed. The stream-callable regression above fails as described.