feat(server): Send metrics via the global endpoint

[jan-auer]

Introduces an option http.global_metrics. When enabled and not in
processing mode, Relay sends metrics to the global batch endpoint at
/api/0/relays/metrics/ instead of envelopes. This endpoint allows for
batched submission of metrics from multiple projects, which should
reduce the overall number of requests.

Bug Fixes

This change contains additional bug fixes that were discoverd during
implementation:

• The batch metrics endpoint no longer requires the emit_outcomes
  outcomes flag to be set. This was invalid copy & paste from the
  outcomes endpoint.
• Requests with HTTP encoding received a signature computed from the
  compressed body. However, Relay requires the signature on the
  uncompressed body.

Details

Building the request occurs in the EnvelopeProcessor in place of
building envelopes in the following steps:

1. Iterate buckets from all projects and stream them into partitions.
2. While collecting, check the size of the partition. As soon as a
   partition reaches the batch size limit, flush the partition eagerly.
   Buckets at the border may be split.
3. At the end flush all remaining partitions with data.
4. To flush a partition, JSON-encode the list of buckets and optionally
   apply HTTP encoding (compression).
5. Submit the SendMetricsRequest with the payload and outcome
   metadata directly to the upstream.
6. In the response callback, log outcomes directly. Therefore, the
   request does not have to be awaited.

In processing mode, Relay still produces to Kafka. The configuration
option does not have an effect in processing mode.

A note on stability: This endpoint and functionality is meant for
operation at scale within a distributed Sentry installation. At this
moment, it is not recommended to enable this option for external Relays.

Tasks

• PR description
• Changelog entry with option
• Integration test
• Doc comments
• BUG: Batch metrics endpoint checks outcomes flag
• BUG: Invalid signature applied to compressed requests

Possible Improvements

The changes in this PR allow for further optimization and changes. In
order to keep the scope of this PR to a necessary minimum, these have
not been included yet:

• Submit regular envelopes directly from the EnvelopeProcessor,
  avoiding the roundtrip through the EnvelopeManager. This also avoids
  a redundant call back to compress envelopes in the processor, should
  HTTP encoding be enabled.
• Improve handling of the ExtractionMode. It can be accessed more
  conveniently through a getter. Additionally, it resides on project
  config where it should actually be included in global config.
  Currently it is being passed around in excess.

[jan-auer]

BucketKeyRef is no longer required. The hash created here and the hash used for partitioning do not have to be the same, which is why we can hash differently.

[jan-auer]

As partitioning is now exclusively in the processor, this metric has moved here from relay-metrics. Its key is still the same.

[jan-auer]

The check for whether the bucket can split has been moved into split_at. Invariants for splitting are still checked in the inner iterator.

[jjbayer]

See comment about test coverage, apart from that LGTM!

[jjbayer]

There's a lot of metrics-specific methods here now, maybe we should move them to a submodule (as standalone functions) in a follow-up PR.