corpus-gen: support `starknet_estimateFee`, `starknet_simulateTransactions` and `starknet_call`

[infrmtcs]

bench/rpc/cmd/corpus-gen covers the read and trace APIs but not the execution methods (see bench/rpc/README.md). Add estimateFee, simulateTransactions and call subcommands so the k6 harness can benchmark them.

Idea

Sample a block N, then use real transactions from block N+1 as the request payload against block_id = N:

• estimateFee / simulateTransactions: request / transactions = a prefix of block N+1's transactions (a contiguous prefix keeps nonces valid and signatures verifiable, so SKIP_VALIDATE is not required). Both take the same broadcasted-transaction list, so they should share one sampler.
• call: derive a FUNCTION_CALL from an invoke transaction in block N+1.

Notes for the implementer

• v0_10 BroadcastedTransaction only accepts version 0x3. L1_HANDLER cannot be broadcast. DECLARE needs the full contract_class instead of class_hash. transaction_hash should be stripped.
• --block-end must be capped at latest−1 so N+1 exists.
• run.js counts JSON-RPC errors as failures, so requests that revert or fail estimation degrade the benchmark. Consider how to keep the corpus clean.

Follow the existing sampler pattern (newSampledCmd, blockRangeFlags, --batch) and update README.md and corpus/all.json.