Consider splitting C419 for sum/min/max into its own rule

[carljm]

As discussed in #3259 (comment), the performance impact of switching from comprehension to generator is different if you are passing to a short-circuiting function (e.g. any or all) vs a non-short-circuiting one (sum, min, max). Given these differing tradeoffs, it would be reasonable for a project to want to enable the lint for short-circuiting cases, and not the others. This suggests that we should split them into separate rules (or add a config option), as requested in #10759 (comment)

[hauntsaninja]

I would also like these to be separate rules, because of the performance impact!

[Andrej730]

Also came across this trying out rule updates from --preview.
Generator is catching up but only when there are a lot of objects to hold in memory (>100k), otherwise list comprehensions are significantly faster.

import timeit

def bench(repeats):
    def decorator(func):
        total = timeit.timeit(func, number=repeats)
        avg = total / repeats
        print(f"{func.__name__: <35} - {avg * 1e6:.2f} μs")
        return func
    return decorator

@bench(repeats=10000)
def test_sum_list_comprehension_4000():
    sum([i for i in range(1, 4000)])

@bench(repeats=10000)
def test_sum_generator_4000():
    sum(i for i in range(1, 4000))

@bench(repeats=100)
def test_sum_list_comprehension_100000():
    sum([i for i in range(1, 100000)])

@bench(repeats=100)
def test_sum_generator_100000():
    sum(i for i in range(1, 100000))

# test_sum_list_comprehension_4000    - 63.88 μs
# test_sum_generator_4000             - 107.81 μs
# test_sum_list_comprehension_100000  - 2700.70 μs
# test_sum_generator_100000           - 3106.45 μs