[metrics] Count per-bucket errors in TabletServer response metrics #4359

Search before asking

• I searched in the issues and found nothing similar.

Description

Motivation

The per-error metrics added in #4213 are updated only by sendError(). Batch APIs such as produceLog, fetchLog, putKv, and lookup can return normal RPC responses containing per-bucket errors, so these failures are not counted.

For example, two buckets returning NOT_LEADER_OR_FOLLOWER should add two events to that error's meter, even when the RPC response completes normally.

Solution

• Generate recursive error-code collection for protobuf messages and update per-error metrics from the normal response path, following Kafka's response error-counting approach.
• Count each failed response entry, excluding absent error codes and NONE. Preserve the existing aggregate metric's whole-request failure semantics and document that per-error metrics need not sum to the aggregate.
• Cover mixed-success batches, repeated errors, nested responses, message reuse, and encoding failures. Metric failures must not trigger another response.

Anything else?

Follow-up to #4213 (review).

Willingness to contribute

• I'm willing to submit a PR!