feat(metering): cache pending trie nodes for faster state root calculation

[niran]

Summary

Cache trie nodes computed during pending flashblock state root calculations and reuse them across bundle metering simulations. This optimization ensures that each bundle simulation only pays the I/O cost for its own state changes, not for all previous flashblock state changes.

Problem

Without caching, every bundle simulation must:

1. Compute the hashed post-state for all flashblock changes
2. Read trie nodes from disk for all accounts/storage touched by flashblocks
3. Then compute the bundle's incremental changes on top

With many bundles being simulated per flashblock, this redundant I/O becomes a bottleneck.

Solution

Introduce PendingTrieCache that:

1. Caches TrieUpdates and HashedPostState from pending state root computation
2. Uses arc_swap::ArcSwap for lock-free reads and atomic updates
3. Provides ensure_cached() for lazy computation and caching
4. Prepends cached data to bundle's TrieInput via prepend_cached()

Changes

• trie_cache.rs: New module implementing PendingTrieCache with ensure_cached() method that computes and caches the pending trie input, or returns the cached entry if it matches the current flashblock.
• meter.rs: Add PendingTrieInput struct containing TrieUpdates and HashedPostState. Update meter_bundle to use cached trie input when available. Use TrieInput::prepend_cached() to prepend pending trie for incremental state root calculation.
• rpc.rs: Integrate with PendingTrieCache to compute/retrieve cached trie data for bundle simulations.
• lib.rs: Export PendingTrieCache and PendingTrieInput.
• Cargo.toml: Add arc-swap and reth-trie-common dependencies.

Performance Impact

Before: Each bundle simulation reads trie nodes for all flashblock changes
After: Each bundle simulation only reads trie nodes for its own changes

The improvement scales with:

• Number of accounts/storage slots modified by flashblocks
• Number of bundles simulated per flashblock

Test Plan

• Existing metering tests pass (20 tests)
• State root calculation produces correct results with cached data
• cargo +nightly clippy passes
• cargo +nightly fmt passes