feat(bench): add independent payment and general pending limits

[decofe]

Summary

Add opt-in --max-pending-payments and --max-pending-general budgets so an accumulated general backlog cannot consume the payment submission budget. Existing shared-limit behavior remains the default.

The experiment supports independent Tempo expiring-nonce transactions only. It uses is_payment_v2 for signed and deferred requests, omits offers to a full class before submission while continuing to feed the other class, and pauses the bounded input stream when both budgets are full. Inclusion, definite rejection, and chain-observed expiry release only the matching class's slot. Per-class pending and throttled-offer metrics make this behavior observable; setup is not classified or omitted.

Profile

The existing public-mix benchmark still shows a throughput dip with a 10k general pool cap because transactions evicted from the node remain outstanding in txgen until inclusion or expiry. This tests sender-side isolation without changing the node or nominal workload mix.

Verification

• New tests cover payment progress with a full general budget, class-specific refill, rejection of inclusion-ordered workloads, flag validation, and signed/deferred classification parity.
• cargo +nightly clippy -p bench-core -p bench-cli -p txgen-tempo --all-targets -- -D warnings passes.
• cargo test -p bench-core -p bench-cli -p txgen-tempo --lib --bins --tests -- --skip replays_a_mixed_corpus_identically_twice passes (one pre-existing ignored test); formatting and the per-class metric test also pass. The excluded Anvil RPC-replay test fails the same response-kind assertion on the unchanged base revision.
• Matched benchmark is running: shared 50k versus 40k payment / 10k general, same Tempo revision and txgen binary, public-mix, 90 seconds, 10 validators across five GCP regions, 100 GiB state, 50k target TPS, Samply enabled, sccache enabled, and VictoriaMetrics export disabled. The harness checks effective sender report metadata and rejects incorrect limits; a mock-sender regression test verifies the actual arguments and rejects omitted flags. Results pending; no performance claim yet.
• The previous run completed but is not a valid comparison: both phases used shared 50k backpressure because the intended split flags did not reach the sender.

Prompted by: @klkvr