aorta.race.modes: PEP 562 `__getattr__` with no `__dir__`, so `__all__` names are missing from `dir()`

[amd-vivekag]

Found by a sweep while addressing Copilot's __dir__ comment on #428, which
raised the same class of issue against aorta/__init__.py. Filed rather than
folded into #428 because src/aorta/race/ is outside that PR's diff.

src/aorta/race/modes/__init__.py defers three exports behind a module-level
__getattr__ (PEP 562) and has no __dir__:

# Lazy imports to avoid circular dependencies
def __getattr__(name: str):
    if name == "DefaultModeReproducer":
        from .default import DefaultModeReproducer
        return DefaultModeReproducer
    ...

__all__ = [
    "create_reproducer",
    "DefaultModeReproducer",
    "DDPModeReproducer",
    "FSDPModeReproducer",
]

So on a fresh import set(__all__) <= set(dir(aorta.race.modes)) is false:
three of the four promised names are absent from dir() until something has
already read them. Tab-completion and help() do not advertise them, which is
the introspection regression Copilot described on #428.

The repo already has an established form for this, in four modules:

def __dir__() -> list[str]:
    return sorted(set(globals()) | set(__all__))

• src/aorta/__init__.py (added in perf(cli): load command groups lazily, cutting aorta --help from 276 ms to 106 ms #428)
• src/aorta/report/__init__.py
• src/aorta/report/analysis/__init__.py
• src/aorta/report/generators/__init__.py

aorta/race/modes is the only module-level __getattr__ in src/ still
missing it.

What would close this

1. Add the same __dir__ to src/aorta/race/modes/__init__.py.
2. A test asserting set(__all__) <= set(dir(...)) on a fresh import in a
   subprocess — in-process the pytest session has usually already resolved
   the names, so the assertion passes vacuously. tests/test_package_version.py
   has the pattern.

Low priority: this is introspection only, no import or runtime behaviour
changes.