perf(cold-start): eth_createAccessList two-shot warming for fast first boot

[KaiCode2]

Stacked on #31 (perf/curve-cold-start) — review/merge that first.

Problem

#31 made repeat warms of layout-free pools (Curve/Balancer) fast, but the
first-ever boot of an unknown pool still ran the get_dy / getPoolTokens
discover call in local revm over a cold cache — faulting each SLOAD one-at-a-time
over RPC. That serial faulting dominates a cold first boot.

Change

Two-shot cold warming: derive the read-set with one eth_createAccessList,
bulk-load it through the EvmCache, then run the discover warm (no serial
faulting).

• Generic cache primitive moved upstream: evm-fork-cache now has a
  cache-owned access-list read-set prewarm primitive on branch
  feat/access-list-read-set-prewarm, commit
  0c9af358d90c87e6452d5b3f1e252822043269f2. This branch is pinned to that
  exact commit until the cache crate publishes the primitive.
• cold_start_many primes fallback pools whose planner declares a discover
  call — generic over Curve/Balancer and future discover-based adapters, with
  zero new adapter trait surface. The per-pool cold-start then runs warm and
  finalizes as usual.
• AdapterRegistry::cold_start_primed(pool, cache, policy) is the single-pool
  async entry point. It no longer asks users to pass a separate provider/RPC; it
  uses the provider/fetchers already installed on the EvmCache.
• Automatic + graceful: default-on, no new API on the happy path. A provider
  without eth_createAccessList, a reverted call, or any per-pool failure
  transparently falls back to local discovery. Opt out with
  with_access_list_discovery(false).
• No correctness risk: the access list only prewarms; the local discover
  stays authoritative and self-heals an incomplete list.

Measured

Paid endpoint rerun requested in review:

set -a; source /Users/kaialdag/Desktop/dev/egg/crates/evm-amm-state/.env; set +a
CURVE_PHASES_ITERS=5 cargo run --release --example curve_cold_start_phases

Run used the paid Alchemy mainnet E2E_RPC_URL from .env with the example's
gzip-enabled HTTP client. Tricrypto2, block 25_481_590, 5 iterations:

• discovery cold_start: 717.8 ms median (min..max 608.1..792.6)
• access-list first boot: 483.6 ms median (min..max 396.4..933.5), 1.5× faster than local discovery
• verify-only cold_start: 116.9 ms median (min..max 96.4..132.7)
• cold_start_many known read-set: 110.7 ms median (min..max 75.3..142.2)

The benchmark section in docs/benchmarks.md has been refreshed with these
paid-RPC numbers.

Interop / API notes

• The gasPrice = pinned block baseFee rule and null-tolerant
  eth_createAccessList response decoding now live in evm-fork-cache, not in
  this AMM crate.
• The local serde dependency added only for lenient access-list decoding was
  removed.
• cold_start_many still accepts a provider because its known-read-set path runs
  explicit storage programs; the new access-list discovery path itself uses the
  cache-owned provider/fetcher primitive.

Verification

• cargo fmt --check
• cargo check --all-targets
• cargo clippy --all-targets --no-deps -- -D warnings
• cargo test --test cold_start_adoption
• set -a; source /Users/kaialdag/Desktop/dev/egg/crates/evm-amm-state/.env; set +a; cargo test --test access_list_discovery_rpc -- --ignored --nocapture
• set -a; source /Users/kaialdag/Desktop/dev/egg/crates/evm-amm-state/.env; set +a; CURVE_PHASES_ITERS=5 cargo run --release --example curve_cold_start_phases