Infinite recursion in translatables unpacking code

I'm not sure if this patch fixes the problem correctly, but it highlights the problematic line.

The gist of the issue is that $storage is passed down in recursion as-is. This (along with a special child key present also in the parent storage) might have caused the algorithm to keep hitting the same part of the array.

I'm not sure if the following is the option_definition triggering the problem. It's the best I can do considering each request is dumping 20000 rows into watchdog every second.

Anyway, here goes:

Array
(
    [id] => date_filter
    [table] => node
    [field] => date_filter
    [relationship] => none
    [group_type] => group
    [ui_name] => 
    [operator] => >=
    [value] => Array
        (
            [min] => Array
                (
                    [date] => 
                )

            [max] => Array
                (
                    [date] => 
                )

            [value] => Array
                (
                    [date] => 
                )

        )

    [group] => 0
    [exposed] => 
    [expose] => Array
        (
            [operator] => 
            [label] => 
            [use_operator] => 0
            [identifier] => 
            [optional] => 1
            [remember] => 0
            [single] => 1
        )

    [date_fields] => Array
        (
            [node_data_field_event_start.field_event_start_value] => node_data_field_event_start.field_event_start_value
        )

    [date_method] => OR
    [granularity] => day
    [form_type] => date_text
    [default_date] => today
    [default_to_date] => 
    [year_range] => -3:+3
)

The infinite recursion is happening in [value].

However, regardless of the problematic array structure, the code as it currently stands doesn't recurse down into the array properly anyway. For example, the [min] and [max] keys are merely skipped because the $storage array passed down simply doesn't contain those keys (and even if it did, the function wouldn't be crawling under the right trail). I suspect any value below the first level are never actually scanned.

P.S. The only reason I haven't flagged this as critical is that it's for translatables. However, I think it deserves attention because it causes out-of-memory error for us, even at 256M.