Retry writes that fail with EOF

[HarveyBase]

Description

The Writer currently treats io.ErrUnexpectedEOF and several connection errors as transient, but a plain io.EOF returned while a broker connection is being closed is not retried. During routine broker rolling maintenance, this can turn a recoverable connection interruption into a permanent write failure.

This change classifies io.EOF as a transient network error, allowing the existing Writer retry loop to handle it without changing retry limits or backoff behavior.

Testing

• Added TestIsTransientNetworkErrorEOF to cover error classification.
• Added TestWriterRetriesEOF using a fake RoundTripper: metadata succeeds, the first Produce returns io.EOF, and the second succeeds. The test asserts the write succeeds after exactly two Produce attempts.
• gofmt and git diff --check pass.
• go test -mod=mod . -run '^TestWriterRetriesEOF$' -count=1 -race -timeout=20s passes.
• go test -mod=mod . -run 'Test(Error|IsTransientNetworkErrorEOF)$' -count=1 passes.
• go vet ./... passes.
• The integration tests requiring Kafka were not run because no local broker was available; the failures are environment connection-refused errors, not code failures.

Fixes #1352