feat(anvil): support multiple fork URLs with round-robin load balancing

[decofe]

Summary

Allow --fork-url to accept multiple RPC endpoints. Requests are distributed across endpoints using a custom round-robin tower::Service — each request goes to the next transport in rotation. On failure, the retry layer above handles retries (which will hit the next transport in the ring).

Closes #14265

Usage

CLI: multiple --fork-url flags

anvil --fork-url https://rpc1.example.com --fork-url https://rpc2.example.com --fork-url https://rpc3.example.com

foundry.toml: endpoints array (backwards compatible)

[rpc_endpoints]
mainnet = { endpoints = ["https://rpc1.example.com", "https://rpc2.example.com"], retries = 5, retry_backoff = 1000 }

Then use the alias:

anvil --fork-url mainnet

Existing single-endpoint configs continue to work unchanged:

[rpc_endpoints]
mainnet = "https://rpc1.example.com"
mainnet_alt = { endpoint = "https://rpc2.example.com", retries = 5 }

Changes

• crates/config/src/endpoints.rs: Added extra_endpoints to RpcEndpoint, support endpoints array in deser (backwards-compatible with existing endpoint key), ResolvedRpcEndpoint::all_urls() helper
• crates/anvil/src/cmd.rs: Changed fork_url from Option<ForkUrl> to Vec<ForkUrl>, alias resolution expands multi-endpoint configs, validates no block suffixes on secondary URLs
• crates/anvil/src/config.rs: Added fork_urls: Vec<String> to NodeConfig, branched setup_fork_db_config to build a round-robin provider when multiple URLs are present
• crates/common/src/provider/mod.rs: Added RoundRobinService (a tower::Service that rotates requests across multiple RuntimeTransports) and ProviderBuilder::build_fallback() which wraps it with a retry layer
• crates/anvil/src/eth/backend/fork.rs: Added fork_urls to ClientForkConfig, updated update_url to rebuild the provider and keep URLs in sync

Design

Uses a custom RoundRobinService that distributes requests sequentially across transports using AtomicUsize rotation. Each request goes to exactly one transport. On error, failover is handled by the RetryBackoffLayer above, which will naturally hit the next transport on retry since the index advances.

Note: There is no health scoring, endpoint quarantine, or active deprioritization of unhealthy endpoints. A flaky endpoint will continue to receive its share (~1/N) of requests. True health-based fallback (e.g. via Alloy's FallbackService) could be added in a follow-up.

Transparent to all upstream consumers — ClientFork, SharedBackend, BackendHandler, ForkedDatabase see the same Arc<RetryProvider> / dyn Provider<N>.

Backwards Compatibility

• Single --fork-url works exactly the same
• foundry.toml single endpoint key unchanged
• endpoints (array) is a new optional key — cannot be combined with endpoint

Known Limitations

• No health scoring or endpoint deprioritization — all endpoints receive equal traffic regardless of error rate or latency
• No startup validation that all endpoints return the same eth_chainId or can serve the fork block
• anvil_setRpcUrl resets to a single endpoint; multi-endpoint state is startup-only
• RoundRobinService::poll_ready() always returns Ready without polling inner transports

Testing

cargo test -p foundry-config -- endpoints
cargo test -p anvil --lib cmd::tests

Prompted by: zerosnacks

[stevencartavia]

setup_fork_db_config clones self.fork_urls into ClientForkConfig, but in the non-fork-to-fork reset path node_config.fork_urls is still the default empty vec. This panics on fork_urls[0] in eth_rpc_url().

[mablr]

Some comments, looks good overall.

[mablr]

sgtm

[zerosnacks]

Made some small bug fix changes here: 8cd02a5 and here: #14393

PR lgtm