fix(tavus): don't retry 4xx when calling the Tavus API, and keep the provider's error

[sidxh]

TavusAPI._post retries every failure, including 4xx responses that can never succeed. With a wrong API key the avatar takes about 4 seconds to fail, logs three tracebacks, and ends with a retryable APIConnectionError("Failed to call Tavus API after all retries") that has lost the 401 and the response body.

The loop catches Exception, which includes the APIStatusError it raises itself for a non-2xx response, logs each one with logger.exception, and always finishes with a generic, unchained APIConnectionError. It also runs range(max_retry), one attempt fewer than the other avatar plugins.

This is the same change #7314 made for Anam. A non-retryable APIStatusError is re-raised immediately. 5xx and network errors are retried max_retry times after the first attempt with a warning (status code only, no payload), and the final error is the provider's APIStatusError, or an APIConnectionError chained to the network error.

Against a local server with default connect options, on main: a 401 takes 4.00s, three tracebacks, generic APIConnectionError; a persistent 503 is the same. With this change a 401 raises APIStatusError(status_code=401) on the first attempt with no traceback, 503, 503, 200 still succeeds, and a persistent 503 raises APIStatusError(status_code=503) after four attempts (6.01s, one more attempt than before, matching Anam). Five tests added in a new tests/test_plugin_tavus.py; the four retry-classification ones fail on main, and one covers a 2xx with a malformed JSON body, which now fails once with a non-retryable APIConnectionError instead of re-sending a POST Tavus already accepted.

Closes #7493.