The gateway's metrics exporter sets the temporality to Delta in
pg_documentdb_gw/documentdb_gateway_otel/src/lib.rs (.with_temporality(Temporality::Delta)),
with no way to change it.

Many OpenTelemetry pipelines that end in Prometheus-compatible storage expect cumulative
sums and histograms. For example, Grafana Alloy's otelcol.exporter.prometheus drops
delta points, and the delta-to-cumulative processor is still marked experimental there.
With the current gateway, the metrics never reach the store unless an experimental
component is enabled.

Request: honour the standard OTEL_EXPORTER_OTLP_METRICS_TEMPORALITY_PREFERENCE
environment variable (cumulative | delta | lowmemory), and/or a
TelemetryOptions.Metrics.Temporality setting in SetupConfiguration.json. The default can
stay Delta.

Two related details we noticed while wiring this up (happy to open separate issues):

1. db.client.operation.duration is recorded in seconds but uses the SDK's default
   histogram boundaries (0, 5, 10, 25, … 10000), so nearly every request falls in the first
   bucket. Second-scale boundaries (e.g. 0.001 … 10) would make the percentiles usable.
2. The unit suffix varies between series of the same metric (some are exported with the
   unit s/By, some without), which gives two Prometheus names for one metric
   (db_client_operation_duration_seconds_* and db_client_operation_duration_*).

Seen at commit cdf2ee2 (1.0 RC1).