Forward reduce_batch_size_fn in find_executable_batch_size decorator factory

[icn5381]

Bug

find_executable_batch_size's parenthesized decorator form drops reduce_batch_size_fn when building its functools.partial, so a custom batch-size reducer is silently ignored and the default multiply-by-0.9 behavior always runs:

@find_executable_batch_size(starting_batch_size=128, reduce_batch_size_fn=my_reducer)
def train(batch_size):
    if batch_size > 10:
        raise RuntimeError("CUDA out of memory.")
    return batch_size

train()          # → 9, after iterating *0.9 twenty-odd times
# my_reducer was never called

Since the first positional parameter is the function itself, the factory form is the only way to pass a custom reducer — the parameter added in #3071 has been unusable, and it is documented in the docstring.

Fix

The functools.partial now forwards reduce_batch_size_fn alongside starting_batch_size.

Verification

• New test_custom_reduce_batch_size_fn asserts the custom reducer is called and its return value is used: fails on main, passes with the fix.
• tests/test_memory_utils.py: 8 passed.