refactor(prover-node): address review feedback on checkpoint-store redesign

[PhilWindle]

Addresses the review feedback on the checkpoint-store redesign (PR #23552). Branches off pw/checkpoint-store-redesign so it can be folded back in.

prover-node

• Advance the local tips store only after block-stream handling succeeds, and propagate checkpoint registration failures instead of swallowing them, so a transient handler failure leaves the tips unadvanced and the L2BlockStream re-emits the event (A-1041).
• Skip checkpoints for epochs already past their proof-submission window (avoids archiver catch-up noise).
• Bound the publishing-service shutdown with a timeout so an in-flight L1 tx can't hang stop().
• expireEpoch fetches blocks via getBlocks({ epoch }) instead of computing the range from getCheckpointsData.

session-manager

• archiverFullyCovered compares by content-addressed id (number, slot, archive root) rather than checkpoint number, so a reorg that keeps the number but changes content is detected.
• openPartialSession reuses a live partial session whose checkpoint content already matches the canonical set instead of reconstructing one.
• startProof refuses to re-prove an epoch the L1 proven chain already encompasses.
• Collapsed the redundant proving/finalization delays into a single delay (they gated identical work).
• startProof returns the job id without awaiting completion; callers poll getJobs() (proving can outlast an HTTP request). RPC output changes void → string.

proof publishing

• Wire the candidate's submission-window deadline into l1-tx-utils (txTimeoutAt) so the L1 tx stops retrying past the deadline.
• Drop the now-redundant waitUntilStartBuildsOnProven loop from ProverNodePublisher — the publishing service already gates publication on the predecessor being proven.

orchestrator

• Abort the sub-tree when the chonk verifier proof fails (otherwise the checkpoint/epoch orchestrators hang forever).
• Cancel the proving state on sub-tree cancel, matching TopTreeOrchestrator.
• Name BaseRollupHintsWithoutProofAndVK at the prepareBaseRollupInputs boundary and correct the stale "mock proof and VK" comment.
• Clarify the proving-scheduler throttle docs: the shared SerialQueue paces job initiation, it does not bound in-flight broker concurrency.

e2e

• Anchor the mid-epoch reorg test on a fresh epoch and add count-based assertions so it provably exercises in-epoch checkpoint removal: N checkpoints in the epoch, remove the last leaving N-1, prove up to and including the (N-1)th.

Prover-node unit tests (162) and prover-client orchestrator tests pass; build and lint clean.