Sessions as sqlite databases

In the beginning, session files used to be pickle. This proved to be bad as soon as one wanted to add more fields. For this reason, they were migrated to use JSON instead. But this proved to be bad as soon as one wanted to save things like entities (usernames, their ID and hash), so now it properly uses sqlite3, which has been well tested, to save the session files! Calling .get_input_entity using a username no longer will need to fetch it first, so it's really 0 calls again. Calling .get_entity will always fetch the most up to date version.

Furthermore, nearly everything has been documented, thus preparing the library for Read the Docs (although there are a few things missing I'd like to polish first), and the logging are now better placed.

Breaking changes

• .get_dialogs() now returns a single list instead a tuple consisting of a custom class that should make everything easier to work with.
• .get_message_history() also returns a single list instead a tuple, with the Message instances modified to make them more convenient.

Both lists have a .total attribute so you can still know how many dialogs/messages are in total.

New stuff

• The mentioned use of sqlite3 for the session file.
• .get_entity() now supports lists too, and it will make as little API calls as possible if you feed it InputPeer types. Usernames will always be resolved, since they may have changed.
• .set_proxy() method, to avoid having to create a new TelegramClient.
• More date types supported to represent a date parameter.

Bug fixes

• Empty strings weren't working when they were a flag parameter (e.g., setting no last name).
• Fix invalid assertion regarding flag parameters as well.
• Avoid joining the background thread on disconnect, as it would be None due to a race condition.
• Correctly handle None dates when downloading media.
• .download_profile_photo was failing for some channels.
• .download_media wasn't handling Photo.

Internal changes

• date was being serialized as local date, but that was wrong.
• date was being represented as a float instead of an int.
• .tl parser wasn't stripping inline comments.
• Removed some redundant checks on update_state.py.
• Use a synchronized queue instead a hand crafted version.
• Use signed integers consistently (e.g. salt).
• Always read the corresponding TLObject from API responses, except for some special cases still.
• A few more except low level to correctly wrap errors.
• More accurate exception types.
• invokeWithLayer(initConnection(X)) now wraps every first request after .connect().

As always, report if you have issues with some of the changes!