Split benchmark fields per type: schema 2 rejects fields the type does not use

[ishandhanani]

Stacked on #404 (Track 2, the deferred benchmark union from #385). Draft until the stack below it merges.

What

BenchmarkConfig is one flat dataclass with every type's fields on it, so a field the runner never reads (isl on gsm8k, num_shots on sa-bench) loads fine and silently does nothing. Each runner now declares the fields it reads as config_fields, next to a shared set every type may set (client_placement, client_dedicated_node, colocate_with_frontend, sweep, aiperf_package, aiperf_args, export_node_metrics).

• A schema: 2 recipe that sets a field outside shared + its type's fields is rejected at load, naming the stray fields and the accepted ones.
• A schema 1 recipe gets a warning and keeps loading, so nothing downstream breaks before it opts in.
• Adding a field to a runner means adding it to that runner's config_fields; a test asserts every declared field exists on BenchmarkConfig.

Checked against the whole corpus

The rule was run over the 555 historical in-repo recipes and the 481 downstream InferenceMAX recipes, 1179 variants after override expansion, each forced to schema: 2. It rejects exactly one: a manual-type recipe carrying isl/osl/concurrencies. Every other field set per type is accepted. (The same pass surfaced two pre-existing, unrelated load failures in the downstream corpus: 98 recipes set telemetry.provider / default_frequency, which the current TelemetryConfig does not have, and 6 set benchmark.tokenizer_mode. Those fail on main today and are not touched here.)

Validation

• Full suite green (1847 passed); lint, schema-docs drift check, every example validates.
• tests/test_benchmark_fields.py: declared fields exist, each type accepts the field sets its recipes use, shared fields work for every type, schema 2 rejects with the expected message, schema 1 warns, defaults never count as set.

[ishandhanani]

Consolidated into #407, which carries these same commits as one PR against main (this description is reproduced there as one of its parts). Closing to keep review in one place; the branch stays for per-step reference via stack #398.