fix: Retry incomplete HTTP responses

[brightsparc]

Fixes #347.

What changed

• Treat nested hyper::Error values for which is_incomplete_message() is true as retryable.
• Preserve the existing exporter backoff and maximum elapsed-time behavior.
• Add a socket-level regression test that accepts an HTTP request, closes without returning a response, verifies Hyper reports IncompleteMessage, and verifies RetryPolicy retries it.

Why

During a staging ClickHouse load test with Rotel 0.2.2, ten turns were missing and Rotel logged permanent exporter drops with this error shape:

TransportErrorWithMetadata {
    original_error: hyper_util::client::legacy::Error(
        SendRequest,
        hyper::Error(IncompleteMessage),
    ),
    metadata: None,
}

Three drops occurred during the affected run at 2026-07-30 00:50:45Z, 01:10:58Z, and 01:33:58Z. Rotel's retry policy already walks the error source chain, but only recognized timeouts and ConnectError, so this transport failure bypassed the configured retry budget and immediately reached Exporting failed, dropping data.

ClickHouse query and asynchronous-insert logs showed successful inserts continuously through all three windows, with status = Ok, exception_code = 0, and no corresponding ClickHouse warning/error entries. That points to an incomplete response caused by a connection reset rather than malformed telemetry or a broad ClickHouse failure.

Delivery semantics

An incomplete response is an ambiguous commit: the destination might have committed the request before the response was truncated. Retrying therefore provides at-least-once delivery and may produce duplicates unless the destination deduplicates requests. For telemetry export, that is preferable to confirmed permanent loss; stable request identity/deduplication remains a worthwhile follow-up for destinations that need stronger guarantees.

Validation

• cargo fmt --all --check
• cargo test -p rotel exporters::http::retry::tests — 18 passed
• cargo clippy -p rotel --lib --tests -- -D warnings — blocked by existing repository-wide warnings in unrelated files; this change introduces no Clippy diagnostic in the output

[mheffner]

Looks good, thanks for fixing!