feat(sdk): serialize/hydrate LruMemoCache for SSR

[arkhindi12]

Closes #977.

Problem

LruMemoCache lives in memory. In a Next.js server component the instance dies
with the request, so every request repeats the same contract-metadata queries
that the cache was supposed to absorb.

What this adds

serialize() writes a versioned snapshot; hydrate() restores it.

const json = cache.serialize();          // drop into a session/Redis entry
// next request:
cache.hydrate(json);                     // returns entries restored

Also exported as free functions — serializeCache(cache) /
hydrateCache(cache, json) — for callers holding the cache through an
interface.

Details that matter

• Absolute deadlines survive. Hydration restores the original expiresAt
  instead of restarting every TTL, so a snapshot taken at T+9s of a 10s entry
  expires 1s later, not 10s later.
• Expired entries are dropped at both ends.
• LRU order is preserved, so eviction after hydration still drops the
  coldest entry.
• Capacity still applies: a snapshot taken with maxSize: 10 cannot
  overflow a cache built with maxSize: 3.
• Malformed input throws (SyntaxError / TypeError, including a version
  mismatch) rather than leaving a half-filled cache.
• Metrics are not fabricated — a hydrated cache reports 0 hits / 0 misses
  and measuredSpeedupPercent: null, because it has measured nothing.

Tests

src/tests/lru-memo-cache-serialize.test.ts — 10 cases: round-trip, deadline
preservation with fake timers, expired-entry filtering, LRU order after
hydration, capacity clamping, metrics untouched, malformed input rejected,
non-primitive keys rejected, and the exported helpers.

Test Files  1 passed (1)
     Tests  10 passed (10)

Rebased on main after #1004. npx tsc --noEmit and this PR's test suite are green on the current head.