fix(eth-engine): mint a fresh JWT for each retried Engine API request - #394

EngineRpc::rpc_request mints the JWT once, before the request is sent:

.retry(retry_policy)
.bearer_auth(self.auth.generate_token()?)
.send()

RpcRequestBuilder::send then reuses that one string on every retry attempt, and exchange_capabilities retries with ENGINE_EXCHANGE_CAPABILITIES_RETRY_RPC — a constant 3s delay and without_max_times(), so it never gives up.

The Engine API rejects a token whose iat is more than 60 seconds from the server's clock (JWT_MAX_IAT_DIFF in alloy-rpc-types-engine, which is what reth's JwtAuthValidator checks). So if the execution client's auth port takes longer than a minute to come up, the consensus node keeps resending a token the EL can no longer accept, and no amount of further retrying recovers it — the node stays locked out even after the EL is healthy, until it is restarted.

Two attempts 1.1 seconds apart on main carry the byte-identical token:

left:  eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJpYXQiOjE3ODkyMzc4ODksImV4cCI6bnVsbH0...
right: eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJpYXQiOjE3ODkyMzc4ODksImV4cCI6bnVsbH0...

The payload is {"iat":1789237889,"exp":null} in both — the timestamp is frozen at the first attempt.

The builder now takes the Auth instead of a finished token and mints one inside the retry closure. Only EngineRpc sets it; the IPC transport carries no JWT, and every other Engine method uses NoRetry, so nothing else changes.

The added test asserts the two attempts carry different tokens and that both validate against the secret. It fails on main with the output above. cargo test -p arc-eth-engine passes (96), as do fmt and clippy.