…ake (#1256)

- [x] ~Note: stacked on #1255. Merge that PR first.~

## Overview

Extends the in-process fake-intake APM tests (added in #1194) with
**value-level** coverage of trace stats aggregation. The existing stats
tests assert serialization fidelity (a hand-built `ClientStatsPayload`
survives the roundtrip) and stats presence/suppression
(`_dd.compute_stats` truth table). None assert that the *computed*
aggregates reaching the intake are correct.

This PR routes concrete spans through the real pipeline
(`SendingTraceProcessor` → `StatsConcentratorService` → `StatsFlusher` →
fake-intake) and asserts on the decoded `ClientGroupedStats`:

- `e2e_stats_count_aggregates_identical_spans` — 3 identical top-level
spans collapse into one group with `hits == 3`, `top_level_hits == 3`.
- `e2e_stats_counts_errors_separately_from_hits` — `errors` counts only
`error != 0`; `hits` counts all.
- `e2e_stats_sums_span_durations` — `duration` is the nanosecond sum of
aggregated spans.
- `e2e_stats_groups_by_resource` — distinct resources stay in separate
groups.
- `e2e_stats_groups_by_span_kind` — distinct `span.kind` stay in
separate groups.
- `e2e_stats_groups_by_http_status_code` — distinct `http.status_code`
stay in separate groups.

These verify bottlecap's wiring and serialization end-to-end (a span
entering the trace path comes out the intake correctly aggregated),
which libdatadog's concentrator unit tests cannot see. The aggregation
*math* itself is owned and exhaustively tested upstream in
`libdd-trace-stats`.

To support multiple spans, `run_processor_pipeline` was generalized into
`run_processor_pipeline_with_traces(.., traces)`, with a thin
single-span wrapper preserving the existing `_dd.compute_stats` tests
unchanged.

### Out of scope: time-bucketing

Deliberately not added. Bucket assignment lives entirely inside
`SpanConcentrator` and is thoroughly tested in `libdd-trace-stats`
(`test_concentrator_stats_counts` over multiple buckets,
`test_concentrator_oldest_timestamp_cold/hot`, `test_force_flush`).
bottlecap's wiring does not touch bucketing, so a test here would
duplicate upstream coverage and require a production clock seam (the
concentrator seeds `oldest_timestamp` from wall-clock) for no unique
coverage.

## Testing

This PR is only adding more tests. If they pass, we're good.

Ran locally:
- `cargo test --test apm_integration_test`
- `cargo fmt --all -- --check`
- `cargo clippy --workspace --all-targets --features default`

> *"Six tests walk into a bucket. They all collapse into one group —
turns out they were identical the whole time."* — Claude 🤖