[gateway] Implement failed attestation retry with backoff (PRD Section 8)

[yinkscss]

Context

PRD v0.2 Section 8 defines a failed attestation retry policy: max 5 attempts, exponential backoff, no double-pay on retriable RPC errors.

Problem

Gateway submits attestations once; transient Soroban RPC failures are not retried with backoff or dead-letter handling.

Proposed scope

• Retry retriable errors (RPC timeout, fee bump) up to 5 times with 1m/2m/4m/8m/16m backoff
• Do not retry non-retriable errors (InsufficientEscrow, ProgramPaused, missing contributor)
• Check Postgres (program_id, pr_number) and on-chain state before each retry
• Mark dead-letter after max attempts in webhook_events or new attestation_attempts table

Acceptance criteria

• Retriable RPC failure triggers bounded retry without duplicate payout
• Non-retriable errors fail immediately with audit row
• Dead-letter state visible to operators for manual replay ([api] Add admin webhook replay endpoint for failed deliveries #20)

References

• docs/PRD.md Section 8
• crates/gateway/src/attestation.rs