fix: treat improper token properly

[polvalente]

Context

This pull request originated from a Joken crash when sending invalid tokens that contained two ".", like a.a.a or ... These tokens match correctly on [_, _, _] = String.split(token, "."), thus passing a first validation layer (Joken.expand(token)).

Furthermore, when using before_verify or before_validate hooks that used Joken.peek_claims or Joken.peek_header, there could be a problem where said invalid tokens would bleed through as {:ok, :error}, thus causing the crash (as it was expected to return {:ok, %{} = claims} instead of {:ok, atom})

Fix

Instead of relying on just passing through the Base.decode64 error return, it needs to be wrapped as a {:error, improper_token} tuple.

The test added ensures that both scenarios are now returning an error as expected

[victorolinasc]

Can we use the same return as in expand :token_malformed? I think it is easier to who ever uses the library.