Backend-ai

-------------------

- Fix file upload handling in the asyncio version.

- Stringifying exception classes now use the output of "repr()".

------------------

- Improve asyncio exception handling.
Now it does NOT silently swallow CancelledError/TimeoutError and other
non-aiohttp errors!

------------------

- Upgrade to use aiohttp v3 series.

- Improve handling of base directories outside the current working directory
when uploading files for the batch mode execution.

- Display exit code if available in the batch mode execution.

------------------

- Hotfix: Add missing "ai.backend.client.cli.admin" module in the package.

------------------

- Apply authentication to websocket requests as well.

- Fix the client-side validation of client token length.

------------------

- Relicensed to MIT License to motivate integration with commercial/proprietary
software products.

- Lots of CLI improvements!

- Add "terminate" command.

- Add more "run" command options. Now it does NOT terminate the session after
execution by default, and you can force it using "--rm" option.

- Add "admin keypairs" command and its subcommands for managing keypairs.

- Add "admin agents" command to list agent instances.

- "ps" and "admin session" commands now correctly show the client-given session ID
token instead of the master kernel ID of the session.

- Add "logs" command.

- Fix a continuation bug of the "run" command when using the batch-mode, which
has caused a mismatch of run ID management of the agent and the internal task
queue of the kernel runner, resulting an indefinite hang up with two legitimate
subsequent requesting of batch-mode executions.

As being a reference implementation of the execution loop, all API users are
advised to review and fix their client-side codes.

- Now the client sets a custom User-Agent header value as follows:
"Backend.AI Client for Python X.X.X" where X.X.X is the version.