Network errors are not retried during polling

[tosuke]

Network errors during polling (such as server closing the connection) cause immediate failure without retry.

According to RFC 8628 Section 3.5: https://www.rfc-editor.org/rfc/rfc8628#section-3.5

On encountering a connection timeout, clients MUST unilaterally reduce their polling frequency before retrying. The use of an exponential backoff algorithm to achieve this, such as doubling the polling interval on each such connection timeout, is RECOMMENDED.

Currently, network errors like io.EOF make the entire authentication flow fail, even when these are transient issues that could be resolved by retrying.