Minor bug in `torch.compile` tutorial timing helper

[starpiens]

In the “Demonstrating Speedups” section of the following docs:
https://docs.pytorch.org/tutorials/intermediate/torch_compile_tutorial.html#demonstrating-speedups

the timing helper is described as returning the elapsed time in seconds, but the implementation divides by 1024 instead of 1000:

# Returns the result of running `fn()` and the time it took for `fn()` to run,
# in seconds. We use CUDA events and synchronization for the most accurate
# measurements.
def timed(fn):
    start = torch.cuda.Event(enable_timing=True)
    end = torch.cuda.Event(enable_timing=True)
    start.record()
    result = fn()
    end.record()
    torch.cuda.synchronize()
    return result, start.elapsed_time(end) / 1024

However, torch.cuda.Event.elapsed_time returns the elapsed time in milliseconds, so the last line should be:

    return result, start.elapsed_time(end) / 1000

cc @williamwen42 @msaroufim @anijain2305