feat: compile PromQL fallback subtrees in the physical planner

[zzylol]

Stacked on #484.

Rebuilt into the linear stack on main. Conflict resolutions are recorded in the messages of: "feat(physical): add a PromQL per-series window operator"; "feat(physical): compile PromQL fallback subtrees from typed expressions".

Why

physical_planner::compile accepted a Fallback{QueryExpr} node only when the node itself was supplied as an input. ASAPQuery-backend therefore re-parses the PromQL text into its own operator list and maps each operator to promql_values::compile_* (coverage rows 1, 9, 11, 12). This PR is priority item 1 of docs/develop_docs/physical-compile-coverage.md.

What

• compile lowers a Fallback node from its typed expression. The deployment supplies the raw rows of its one selector at promql_fallback::raw_series_input(node), with the schema from promql_fallback::raw_series(expr). Supplying the node's own ID still means "whole result supplied", for example by an external exact engine. A Fallback that reads no selector, such as vector(1), needs no input.
• New Operator::series_window evaluates each series at the query time, or on a subquery step grid, over the left-open window (t - offset - range, t - offset]. It covers instant selection and rate/increase/delta/{sum,avg,min,max,count}_over_time, reusing the temporal kernels. Only delta is new.
• Shapes compiled by the lowering:
  • selectors with offset
  • by aggregates (a global aggregate over no series returns an empty vector)
  • sort, topk by/limit
  • scalar(), vector(literal)
  • arithmetic with one literal
  • f(sel[R:S]) and f(g(sel[r])[R:S]) subqueries, including subquery offset, capped at 100000 steps
• with_series_identity now accepts these shapes.
• Frontend bug fix: subquery offset/@ were silently dropped. They are now kept as a TimeShift around the subquery.

Before this PR

max_over_time(m[5m:1m]) retained as a Fallback fails to compile: physical operation has no native binding. The same happens for rate(m[5m]), scalar(m) and similar queries. max_over_time(m[5m:1m] offset 1m) lowered to the same tree as the query without the offset.

After this PR

Take the samples m{job="a"} = (50s: 1), (110s: 7), (170s: 3), (290s: 4). The Fallback compiles to SeriesWindow(instant, 60s lookback, steps 60s..300s) → SeriesWindow(max, 5m). At 300s it returns {job="a"} 7. The step at 240s has no sample in its lookback, so it contributes nothing.

Remaining

• Fallbacks with more than one selector, such as vector-vector binaries.
• histogram_quantile: the IR emits by () and drops output labels.
• Subquery operands other than one per-series function.
• Implicit subquery resolution.
• @ modifiers.
• without grouping.
• irate, changes, and other range functions.
• The comparison, summary-algebra and precompute items already listed in the coverage doc.

Validation

• cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings, and cargo test --workspace --no-fail-fast all pass.

• New tests/promql_fallback.rs compares against hand-computed Prometheus results. Cases: counter resets and extrapolation, left-open boundaries, offsets, lookback and stale markers, NaN, empty inputs, scalar() on 0, 1, or 2 series, subquery grid alignment and offset, and the step cap and input contracts.

• Regression tests that fail before their fix: subquery offset retention in promql_lowering.rs, and topk by partitioning.

• A separate reviewer agent that did not write the code reviewed the change. Fixed findings, each with a test:

  • topk by lost its partition.
  • Subquery offset was dropped.
  • The step grid was unbounded.
  • A consumed bare selector could be turned into instant selection.

  Documented rather than changed: output rows keep __name__ in the identity (the adapter applies metric-name rules), and the raw-row extent a subquery requires.

🤖 Generated with Claude Code

[zzylol]

Integration note: #477 moves with_series_identity into asap-types as with_promql_series_identity (so asap-aware-mapping can use it), and promql_rows::with_series_identity becomes a delegate. This PR and #487 extend the physical-layer copy (TimeShift, subqueries, scalar(), literal and vector-vector arithmetic, without). When these land after #477, the extended rules must move into asap-types (with scalar_literal inlined there). The integration/planner-for-backend branch already does this port and passes the full workspace tests.