perf: Cache module dirs

[Zirak]

Calling dir on modules is expensive - in a benchmark, it took up 2% of the time when calling freeze_time, when _get_cached_module_attributes accounted for 4% of the time.

This commit hashes dir calls per module id. This drastically speeds it up. Running with pytest-benchmark:

import time
from datetime import datetime, timezone
from unittest.mock import patch

from freezegun import freeze_time

def setup_with_freezegun() -> None:
    current = datetime(2025, 1, 1, tzinfo=timezone.utc)
    with freeze_time(current):
        time.time()

def setup_with_patch() -> None:
    current = datetime(2025, 1, 1, tzinfo=timezone.utc)
    current_time = current.timestamp()
    with patch("time.time", return_value=current_time):
        time.time()

def test_benchmark_freezegun(benchmark) -> None:
    benchmark(setup_with_freezegun)

def test_benchmark_patch(benchmark) -> None:
    benchmark(setup_with_patch)

We get the following:

----------------------------------------------------------------------------------------------------- benchmark: 4 tests -----------------------------------------------------------------------------------------------------
Name (time in us)                                  Min                   Max                  Mean              StdDev                Median                IQR            Outliers          OPS            Rounds  Iterations
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_benchmark_patch (0001_5f171db)            62.3360 (1.0)      8,936.2020 (14.71)       84.5774 (1.0)      175.7804 (7.06)        70.2175 (1.0)      13.8290 (1.0)         5;839  11,823.4918 (1.0)        4730           1
test_benchmark_patch (0002_5f171db)            62.3370 (1.00)     9,183.8990 (15.11)       86.2260 (1.02)     192.4761 (7.73)        72.3760 (1.03)     14.2655 (1.03)        3;676  11,597.4350 (0.98)       3949           1
test_benchmark_freezegun (0002_5f171db)       376.8070 (6.04)       607.6380 (1.0)        405.6626 (4.80)      24.8866 (1.0)        400.5450 (5.70)     19.8857 (1.44)         10;4   2,465.1026 (0.21)        121           1
test_benchmark_freezegun (0001_5f171db)     1,983.5520 (31.82)    2,392.4130 (3.94)     2,103.8566 (24.87)     74.1705 (2.98)     2,083.2020 (29.67)    69.4740 (5.02)         14;5     475.3176 (0.04)         82           1
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Where 0001 is before this commit, and 0002 is after: Mean time reduced 2,103 ms to 405 ms.

This comes with a tradeoff, reflected in the added tests: Dynamically added attributes will not be picked up. For example:

time_after_start = None

def add_after_start() -> None:
    import time
    import sys

    setattr(sys.modules[__name__], 'dynamic_time_func', time.time)

Because the result of dir is cached, the dynamic_time_func is not picked up.

I leave it up to the maintainers to decide if this tradeoff is worthwhile, considering that there are some other existing blindspots (also shown in added tests).

This commit does not include the benchmark itself - if desired, it can trivially be added.

[Zirak]

Here's the result of pytest-benchmark compare --histogram:

Here's the icicle chart with snakeviz from before the commit:

And after: