fix: reject duplicate label sets after dropping the metric name

[zzylol]

Stacked on #489.

Why

PromQL arithmetic with a literal drops __name__. If two series then have the same label set, Prometheus fails with "vector cannot contain metrics with the same labelset". An independent review of backend #801 found that Planner returned both rows instead.

What

• Operator::series_without_name: a SeriesLabels rewrite that drops __name__ and errors when two rows end up with the same label set.
• Both literal-operand paths use it: per-series Binary over stored readouts (feat: compile per-series vector arithmetic and match Prometheus summation #489), and the Fallback literal path, which previously kept __name__.
• Matching and without relabeling still allow repeated label sets. series_binary already rejects right-hand duplicates and left-hand duplicates that match (feat: compile multi-selector PromQL subtrees and more range functions #487), and without aggregation merges repeats, as Prometheus does.

Before this PR

Series m{job="api",instance="x"} and n{job="api",instance="x"}, evaluated as sum_over_time(m[5m]) * 2, returned two identical {instance="x",job="api"} series. The Fallback a * 2 returned rows that still had __name__="a".

After this PR

Both paths fail with vector cannot contain metrics with the same labelset. Series that stay distinct are unaffected, and the Fallback a * 2 now drops __name__.

Validation

• New regression tests fail before the fix and pass after it: per_series_scalar_arithmetic_rejects_label_sets_equal_without_the_name and literal_arithmetic_drops_the_name_and_rejects_equal_label_sets. A new without case confirms that series equal after dropping the name are grouped, not rejected.
• cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace --no-fail-fast: all pass.

🤖 Generated with Claude Code

[zzylol]

Included in the squash merge of #493 into main (deb3904).