feat(interop): make autorelay retries + backoff configurable, fix dead retry path

[abhicris]

Summary

relayMessageWithRetry was always called with maxRetries=1 at both call sites in interop/relayer.go (today's lines 121, 188), so the exponential backoff inside the retry loop was dead code and any single transient destination-chain RPC failure dropped a cross-chain message permanently. This PR makes the retry behavior real and configurable.

Changes

• Two new CLI flags (also envvars):
  • --interop.autorelay.retries (default 3) — max relay attempts per message before giving up. Set to 1 to reproduce the prior no-retry behavior.
  • --interop.autorelay.retry.backoff.max (default 8s) — caps the 1<<attempt exponential backoff so high retry counts don't trigger unbounded sleeps.
• NewL2ToL2MessageRelayer(logger, maxRetries, retryBackoffMax) signature now accepts the new parameters; orchestrator passes them through from CLIConfig.
• Both call sites (relayMessageWithRetry(..., 1)) now pass r.maxRetries.
• relayMessageWithRetry caps each backoff at retryBackoffMax (when > 0).

The defaults (3 attempts, 8s cap) are conservative — users get improved liveness against transient RPC blips with no surprise; downstream tests that depend on the no-retry behavior can opt out.

Why

Refs #291 ("Supersim autorelay flakes"). The Promise#dispatchCallbacks and waiting-pool retry paths both relied on relayMessageWithRetry to absorb transient errors, but the hardcoded 1 meant the retry loop never iterated. Local repro: any transient eth_sendTransaction failure (port contention, mempool full, restarted Anvil) would drop the cross-chain message and leak the waiting-pool entry.

Compatibility

• NewL2ToL2MessageRelayer is internal — only orchestrator.NewOrchestrator calls it, updated in this PR. No exported API removed.
• Default behavior is now "retry up to 3 times with capped backoff" instead of "no retries". Anyone explicitly relying on no-retry semantics can pass --interop.autorelay.retries=1.
• maxRetries <= 0 is normalized to 1 to keep the loop body running at least once.

Test plan

• go build ./...
• go vet ./...
• go test -run TestL2ToL2Message -count 1 ./interop/... (passes; interop package OK in 0.985s)
• CI

This PR intentionally stays mechanical (config + dead-code revival) and does not also try to fix the InvalidTimestamp race in #291 — that's a separate concern that should ride a dedicated PR so the two changes can be evaluated independently.