Add lightweight consumer metrics collector

## Context

Follow-up to the producer metrics API from #1166 / #1171.

`AIOKafkaConsumer` has no equivalent of `metrics_collector`. Teams that need fetch latency, lag, throughput, or commit timing still have to wrap application code and cannot see fetcher / coordinator internals (#38, #562).

## Proposal

Add `ConsumerMetricsCollector` using the same pattern as `ProducerMetricsCollector`:

- sync callbacks on the hot path
- no aggregation / sampling / backend inside aiokafka
- keyword-only callback arguments
- no-op defaults so subclasses override only what they need
- `metrics_collector=` on `AIOKafkaConsumer`
- experimental until the shape stabilizes

## Initial callback set

| Callback | Intent |
|---|---|
| `on_fetch_completed` | fetch RTT + records/bytes returned (per topic-partition) |
| `on_fetch_failed` | fetch error after retries / terminal failure |
| `on_records_consumed` | records/bytes handed to the application via `getone` / `getmany` |
| `on_records_lag` | highwater − position after a fetch update |
| `on_commit_completed` | offset-commit latency |
| `on_commit_failed` | offset-commit failure |

`topic` and `partition` always passed where applicable; collectors decide on label cardinality. Durations in seconds.

## Out of scope for the first PR

- rebalance / heartbeat lifecycle metrics
- connection / broker-level metrics
- OpenTelemetry helper package

## Implementation notes

- Mirror producer conventions (`Null*` / default no-ops, typo warning for unknown `on_*`, exceptions logged and ignored).
- Likely hook points: `Fetcher` (fetch + lag), consumer poll path (records consumed), group coordinator (commits).
- Docs + example extension of the existing Prometheus example; `CHANGES.rst` entry.

## Questions

1. Is this initial callback set the right first cut?
2. Should lag be reported from the fetcher on every successful fetch, or only when the application consumes?