perf: Avoid array copies and pass only declared callback arguments

[PresentKim]

Summary

Follow-up to #1. Brings the library closer to native speed and adds the script that generates the method traits.

Changes

• No array copy when building results. The userland constructor made ArrayObject copy the whole array, which doubled the cost of every *From() call. Results are now built by cloning an empty instance and calling exchangeArray() (ArrayUtils::blank()).
• Callbacks receive only the arguments they declare (ArrayUtils::callbackArity()). Callbacks that take 1 or 2 arguments use array_map() / array_filter() / array_reduce() or a loop with fewer arguments; callbacks with 3 or more parameters, variadics, or an undeterminable signature keep the old behavior. As a side effect, internal functions such as strtoupper work as callbacks.
• last() / keyLast() use array_key_last() on PHP 7.3+ (the old end() path remains for 7.1/7.2).
• Trait generator: tools/generate-traits.py is now the single source of the traits (--check verifies they are up to date). Generated traits are marked @generated.
• Benchmarks: added the missing @Warmup to FirstLastKeyExistsBench and UniqueReplaceBench. Without it the one-time autoload cost was counted, which inflated results of those subjects (including some numbers in refactor: Replace magic methods with explicit methods split into feature traits #1).
• Tests: callback arity paths, non-closure callbacks, independent result instances, last/keyLast, and that pop() / shift() remove the element from the instance.

Behavior changes

• Callbacks that declare fewer than 3 parameters no longer receive the extra arguments.

Benchmark: main vs this branch

PHPBench suite (1000-element arrays, arrayutils variant, mean per call), both sides with the @Warmup fix, run sequentially on the same machine with PHP 8.3. Not run on PHP 7.x.

	main	this branch
Geometric mean vs native	1.94x	1.51x
Benchmarks within 1.25x of native	23 / 56	41 / 56

Callback-heavy methods (μs per call; native array_map is 23.3, array_reduce-style loop for reduceRight is 25.4):

method	main	branch
map	40.6	24.1
reduceRight	41.6	25.5
reduce	34.7	24.5
every	29.8	22.3
filter	42.3	35.6
forEach	30.6	24.8

What is left

• Wrapping the result in an ArrayObject costs about 0.25μs, or about 1μs when the source array is shared (e.g. values() on a list). This only matters where the native call takes under 0.1μs (values, keyRandom, flat, fill, slice, column). The *As variants return plain arrays and do not pay it.
• On arrays with only a few elements, checking the callback arity (about 0.1μs) can cost more than it saves (flatMap 0.68 → 0.80μs).

Verification

• Every method (instance, As, From, FromAs) was compared against the previous implementation on 8 inputs and about 1,100 calls; results and instance state matched.
• vendor/bin/phpunit: 13 tests, 81 assertions pass (PHP 8.3).
• python3 tools/generate-traits.py --check reports the traits are up to date.