Simplify varray

[YYF233333]

Reduce complexity of these two functions with initializer list constructor.

varray should be ok because it is used together with sarray, so not performace critical.

For vformat I run both microbenchmark and a test project:

code

#define ANKERL_NANOBENCH_IMPLEMENT
#include "nanobench.h"
#include "core/templates/list.h"
#include "core/templates/local_vector.h"

class Test
{
public:
    static void bench()
    {
        String fmt = "Test vformat %s %04d-%02d-%02d %02d:%02d:%02d";
        auto bench = ankerl::nanobench::Bench()
            .minEpochIterations(1000);
        bench.run("vformat", [&]{
            for (int i = 0; i < 100; i++) {
                auto r = vformat(fmt, "test", 2023, 10, 5, 12, 30, 45);
            }
        });
    }
};

result

master:

ns/op	op/s	err%	total	benchmark
223,433.91	4,475.60	0.4%	2.67	vformat

This PR:

ns/op	op/s	err%	total	benchmark
220,315.97	4,538.94	0.4%	2.63	vformat

code

extends Node2D

func _ready() -> void:
	var times = []
	for i in range(10):
		var start = Time.get_ticks_usec()
		for j in range(100000):
			Time.get_datetime_string_from_unix_time(1751215307)
		var end = Time.get_ticks_usec()
		print((end - start) / 1000.0, " ms")
		times.append((end - start) / 1000.0)
	var avg = 0.0
	for t in times:
		avg += t
	avg /= times.size()
	print("average: ", avg, " ms")

result

master:
175.029 ms
172.557 ms
170.201 ms
172.307 ms
172.757 ms
172.753 ms
172.192 ms
171.617 ms
171.543 ms
173.941 ms
average: 172.4897 ms

This PR:
171.95 ms
170.445 ms
170.952 ms
168.081 ms
169.044 ms
170.402 ms
170.3 ms
171.469 ms
171.817 ms
169.767 ms
average: 170.4227 ms

both tests show no regressions.

[Ivorforce]

Makes sense to me to simplify these functions; they're definitely more complex than they need to be.
However, initializer_list has a hidden cost for non-trivial types. You can read about this problem here: godotengine/godot-proposals#12247

We've chosen to ignore this problem in most cases, since it's unlikely to lead to problems for most allocations.
However, as commonly called functions, varray and vformat deserve some extra care.

In practice, instead of using initializer_list, I'd probably create the vector, reserve on it (#105928 is scheduled to be merged early in 4.6), and then push_back items from p_args sequentially. This should both simplify and optimize the code. (or, alternatively, resize and loop through the args, assigning to ptrw(). This should be even faster but might be more annoying to implement).

[YYF233333]

I changed vformat. For varray it is both not performace critical and also I think it is more suitable as a macro in its current state (same for sarray), as long as you don't mind we have a bit more macros.

[YYF233333]

I forget there is already #103917 😅. That should supersede vformat part of this PR. I'll revert them.

[Ivorforce]

I just realized args necessitates a copy anyway, in the same vein that initializer_list does. Therefore, this shouldn't be a performance regression.

The whole initializer_list thing is still relevant, but since it has no satisfying solution yet, I'm OK just simplifying the code for now.