feat: add BuilderTransactions abstraction for end of block builder txs

[MavenRain]

Summary

Closes #343. Adds BuilderTransactions, a trait for the transactions the builder itself appends at the end of the block, with the two phases the issue asks for: validation plus block space reservation before building, and the list of txs to include after building.

Problem

Everything the builder adds for itself is hardcoded today. create_payout_tx has exactly three production call sites (the combined refunds, the proposer payout, the per-bundle kickback) and the space for them is reserved in one place in BlockBuildingHelperFromProvider::new_with_execution_tracer. There is no way for a builder to add a block signature, an attestation or any other tx of its own without editing the finalization path, and no way to make sure the block keeps room for it.

Fix

crates/rbuilder/src/building/builder_tx.rs adds:

• BuilderTransactions: validate and reserve_block_space run once before building, on the state of the parent block; transactions runs at finalization and returns the signed txs to append.
• BuilderTransactionError, named after the equivalent type in op-rbuilder's builder_tx.rs so the two builders read the same way.
• reserve_builder_txs_block_space, which validates every registered source and folds their reservations, passing the running total so a source can fit itself into the remaining budget.

Sources are registered per block with BlockBuildingContext::with_builder_transactions. Carrying them on the context rather than on BlockBuildingHelper follows the existing Arc<dyn RootHasher> precedent and leaves the helper trait, both of its test doubles and every dyn BlockBuildingHelper call site untouched.

Two behaviors are worth calling out:

• The proposer payout tx is not modelled as a BuilderTransactions. It has to stay the last tx of the block (PartialBlock::finalize expects it there, ReceiptsData::pre_payment_logs_bloom is defined as the bloom of everything before it, bid adjustments index len - 1, and backtests infer the fee recipient from the last tx) and its value is only known once the bid is chosen. Builder txs are inserted before it.
• Builder txs are committed in the same window as the combined refunds, that is under !adjust_finalized_block, so a finalization adjustment re-executes only the payout tx and leaves them untouched. Letting a builder tx depend on the payout value would mean generalizing FinalizeRevertStateCurrentIteration to an N-tx tail and teaching the receipts and tx-root caches about it; that is a much larger change and is not attempted here.

The reserved gas also widens the true_block_value deduction, because the bid ceiling is read before finalization and the builder pays for its own txs. With no sources registered, end_of_block_gas == payout_tx_gas and the whole path is a no-op.

Testing

cargo test -p rbuilder --lib builder_tx, covering: only the payout tx lands when nothing is registered; builder txs land before the payout tx with consecutive nonces and the payout stays last; a reseal reverts and rebuilds only the payout tx while the builder txs survive; a tx that cannot fit fails the block; space and value reservations are summed and chained; a failed validation aborts before any reservation. The payout test harness now funds the builder signer and can build an EOA fee recipient, which is what makes committing real end-of-block txs in a unit test possible.

Note for reviewers

Where the validation hook should be used first is #296 / #929 (checking the coinbase can cover the payout before building). I did not move that check in here to keep this PR to the abstraction itself; happy to fold it in as a BuilderTransactions impl in a follow-up, or to rebase this on top of that PR if it lands first.