compute: extend temporal bucketing to Reduce's input arrangement

[ggevay]

fec2af8 introduced temporal bucketing on PlanNode::ArrangeBy: the
lowering picks an ArrangementStrategy from a has_future_updates
analysis bit, and ensure_collections inserts a bucketing operator in
front of the arrangement when the strategy is TemporalBucketing.
TopK, Threshold, and Negate get bucketing transitively, because
their lowering wraps the input in a synthetic ArrangeBy.

Reduce is left out of both hooks: its LIR node has no strategy field,
and its render path (render_reduce via KeyValPlan) bypasses
ensure_collections. As a result, a Temporal Filter -> GroupAggregate
pattern — where no ArrangeBy sits between the temporal MFP and the
reduce — is never bucketed today.

Add a temporal_bucketing_strategy: ArrangementStrategy field to the
Reduce LIR node and set it during lowering from the input's
has_future_updates flag. At render time, render_reduce applies
apply_bucketing_strategy to the (key, val) stream right before
render_reduce_plan arranges it internally.

The lowering also clears LoweredExpr::has_future_updates on bucketing
absorption, so a stack of temporal MFPs only buckets at the lowest
layer; a trailing temporal MFP fused above naturally re-arms the flag.
The render-side bucketed: bool safety net in ensure_collections is
still needed to prevent double-bucketing within a single call (the raw-
collection site vs. the collections.arranged loop, or across multiple
keys in that loop).