removed 'flatten'

[evilbluebeaver]

I encountered an error when I tried to deserialize config

#[derive(Debug, Default, Deserialize)]
struct Config {
    pub pools: HashMap<String, PoolConfig>,
}

#[derive(Debug, Default, Deserialize)]
struct PoolConfig {
    pub endpoints: Vec<String>,
    pub size: usize,
}

using these environment variables

POOLS_FOO_ENDPOINTS="http://localhost:2379"
POOLS_FOO_SIZE = "3"
POOLS_BAR_ENDPOINTS="http://localhost:2379"
POOLS_BAR_SIZE = "3"

It causes an error
Failed to load config: Error(missing field `endpoints`)

I found that the flatten method is causing all the troubles and it is unsound even for loading all the environment variables into HashMap<String, String> because it creates a key for every underscore. So I decided to remove it and replace it with just plain keys method.
That change fixed my issue and also an issue in #49 (I uncommented double_inner_mapping_with_enum_keys and it worked flawlessly)

[evilbluebeaver]

Sorry for closing and reopening pull request. It was a misclick.

[Xuanwo]

I somehow feel like this could be addressed by #56

[Xuanwo]

I'm afraid that this is valid use case we need to support.

[evilbluebeaver]

I doubt that it is possible to support this feature with the same algorithm that is used for structured and specified data.
And also that does not look like valid case for me. Because it creates many misleading keys if an initial key has underscores. For example, if you have a FOO_BAR_BAZ=1 it creates FOO=, FOO_BAR= and FOO_BAR_BAZ=1 which can be misinterpreted by the underlying code.

[evilbluebeaver]

I also added a test for HashMap with complex keys.
I don't know how to implement this correctly where key names aren't known in advance.
So I guess my solution is a half-measure and you feel free to decline it.