feat(query): maintain shared current-series quantile and TopK state

[zzylol]

Rebuilt on top of main (ca944a6a) as the current-series feature alone.

Scope change from the original PR

The window / full-window half of this PR is dropped: #712 landed an equivalent
while this branch was open, and main's version is kept wherever the two overlapped
(covers_range, full_window_slide_ms derivation, the compatibility fixtures).
Rebasing the old branch as-is would have silently reverted #712.

What remains is the part main does not have: CurrentSeries is absent from main
entirely. 43 files, down from 49.

What this adds

Multiple current-value quantiles and TopK limits over the same metric share one
bounded maintained population instead of each planning its own state. Lowering
follows the Planner-selected IR rather than reconstructing intent from query text,
so a catalog rename cannot change the physical plan.

• SeriesPopulation / SeriesReadout contract and LogicalOperator::CurrentSeries
• Backend CurrentSeriesStore: bounded membership, lookback expiry, shared ordered
  state feeding quantile / TopK / sum / count / avg readouts
• Planner bumped to b8b5d705 for typed ExactKind::Min, with the Min readout
  wired through catalog resolution and the exact-agg merge path
• asap_current_series_populations / _cache_builds_total on /metrics

Review fixes carried in

A review of the original branch produced five findings; four are fixed here.

• Budget overflow no longer latches permanently. It records when the budget
  blew and re-arms once the lookback window has moved past it. Previously one
  transient cardinality spike stranded the population on Prometheus until an
  operator forced a replan.
• Quantile readout guards an empty value set instead of underflowing
  values.len() - 1 — that panic would have poisoned the mutex every remote-write
  batch waits on.
• CompileError::Query reports query_id, not the query text.
• Full-window sketch reads skip a layout-provably-empty sid instead of
  discarding the sids already accumulated and dropping the query to the exact path.

The fifth finding (backend install gate missing a full_window_slide_ms check)
was withdrawn on inspection: QueryPlan::validate_against_catalog already
rejects that shape, unconditionally, before the pane/origin loop runs.

Deliberately not included

The original branch also reversed count() over HLL from register-merge to
declined, deleting the ASAPController#163/#165 rationale comments without
mentioning it in any commit message. That is a behavioral change unrelated to
current-series — PromQL count() arguably should count series rather than
report merged cardinality, but it deserves its own PR and its own argument.
Main's semantics are kept here.

Verification

cargo fmt --check, clippy --workspace --all-targets -D warnings, and
cargo test --workspace -- --test-threads=1 — 2139 passed, 0 failed, 4 ignored.

🤖 Generated with Claude Code