runner: resolve benchmarks at runtime, drop the allowlist

[adriangb]

Motivation

Follow-up to #13's review: the bench_standard runner hardcoded an
is_criterion_harness name match to decide how each benchmark was compared,
and the controller maintained a per-repo allowlist (standard / criterion
name lists). Both drift as upstream evolves — e.g. run_tpch is itself
Criterion-harness-based upstream now, and the SQL-harness suite set keeps
growing.

This implements the two asks from that thread:

1. "first try criterion and then fall back to bench.sh so we don't have to match like this"
2. "remove the benchmark allowlist and always try to resolve any benchmarks the user requests"

What changed

Runtime resolution instead of name matching. For each requested
benchmark the runner asks cargo metadata whether the name is a real
Criterion [[bench]] target:

• Yes → cargo bench --bench <name> -- --save-baseline <side>.
• No → bench.sh run <name> (TPC-H keeps its direct dfbench shortcut).

SQL_CARGO_COMMAND is now set unconditionally on the bench.sh path.
bench.sh only reads it inside its SQL-harness run functions, so it's a no-op
for the dfbench suites and gives the harness suites (wide_schema, …) a named
baseline for critcmp — no name list required.

Comparison driven by artifacts, not names: results/*.json →
bench.sh compare_detail; target/criterion baselines → critcmp. A run
mixing both families emits both sections. is_criterion_harness is gone.

Allowlist removed. Any requested name is scheduled and resolved on the
runner (a name that matches neither a Criterion target nor a bench.sh suite
fails there). RepoEntry now carries only kind ("datafusion" | "arrow")
and default_standard. JobType/BenchType collapse Standard+Criterion
into a single Datafusion variant (legacy "standard"/"criterion" strings
still map to it for in-flight DB rows). bench_standard.rs + bench_criterion.rs
merge into bench_datafusion.rs; bench_arrow.rs is unchanged (separate repo).

Deploy note

The BENCHMARK_CONFIG schema changed (services/controller.ts), so this
needs a Pulumi apply on deploy.

Testing

cargo fmt --check, cargo clippy --all-targets -- -D warnings, and
cargo test (127 tests) all pass.

🤖 Generated with Claude Code

[adriangb]

Keeping up the whitelists has been a PITA. Let's try this out.