refactor!: move pane construction out of the physical layer

[zzylol]

Stacked on #475.

Why

Agreed layering: the Planner physical layer compiles computation over typed
input contracts; raw ingestion, pane construction and geometry, storage formats
and stored-state decoding belong to the deployment. The #462 split already moved
stored-state decoding out (#474/#475 have no stored_state), but pane
construction is still in the physical layer: compile_temporal_pane_candidate
decides pane width, count, offsets and entity guards, and the PaneInput
operator checks pane geometry at run time.

What

• Remove physical_planner::temporal_panes (compile_temporal_pane_candidate,
  TemporalPaneMaintenance, TemporalEntityIdentity) and the PaneInput
  operator. ScopeTimestamp stays, moved to operators/scope_timestamp.rs.
• Remove the e2e test that exercised the pane compiler
  (selected_temporal_lifecycle_compiles_panes_and_executes) and its helpers.
• Design doc: one paragraph on what the physical layer does not own.

Computation stays: summary build, pane-state merge (e.g. KLL merge), sketch
estimates and exact finalization. The generic precompute/query split by timing
(#482, #479) is unchanged.

Before this PR

Planner physical layer: lifecycle + window → pane width/count/offsets → PaneInput guards → merge → quantile

After this PR

Deployment: builds panes, binds typed KLL states to input slots
Planner physical layer:                         merge → quantile

Assumptions for review

• Pane merges and sketch estimates are computation and stay in Planner.
• Compiled physical plans keep their own serialized form (plan handoff, not storage).
• The backend takes over pane construction; the removed code and test are saved
  for that follow-up.

Validation

cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings pass;
cargo test --workspace --no-fail-fast: 1355 passed, 0 failed.

🤖 Generated with Claude Code