feat: Make ClusterIndex thread-safe for concurrent lookup

[xiaoxmeng]

Summary:
Make ClusterIndex::lookup() and keyAtRow() safe for concurrent use from multiple threads. Previously crashed with SIGSEGV when called from 20+ benchmark threads.

Three categories of shared mutable state are fixed:

1. Per-call scratch buffers to locals: Remove partitionLookups_ and scratchRanges_ members. lookup() creates local vectors and passes them by reference to helper methods. Zero synchronization overhead.

2. folly::SharedMutex for partition loading: loadPartition() uses std::shared_lock fast path + std::unique_lock with double-checked locking. loadPartitions() (batched preload) takes a write lock.

3. Per-partition folly::SharedMutex for chunk access: getDecodedChunk() uses shared lock for cache hits, exclusive lock only for the slot install (not the IO/decode). Decode happens outside the lock to avoid blocking concurrent readers.

4. Stateless seek() and get() for PrefixEncoding and TrivialEncoding: seek() and get() no longer mutate member state, enabling concurrent calls on the same encoding without locks. PrefixEncoding uses local state via static decodeEntryAt() helper. TrivialEncoding lazily caches string positions via std::call_once.

Benchmark results (index_projector, 60s, --pin_index=true, single shard):

Without preload (--preload_index=false):

Threads	Batches	Throughput	Avg CPU/batch	Peak Mem
1	644	10.50 GB	58.92ms	204 MB
2	2,185	35.29 GB	54.42ms	224 MB
4	4,422	71.48 GB	53.78ms	264 MB
8	8,719	140.75 GB	54.50ms	332 MB
16	16,590	267.69 GB	56.96ms	467 MB
32	25,175	406.29 GB	73.11ms	723 MB

With preload (--preload_index=true):

Threads	Batches	Throughput	Avg CPU/batch	Peak Mem
1	1,097	17.79 GB	54.10ms	195 MB
2	1,719	27.79 GB	57.33ms	212 MB
4	4,393	71.01 GB	54.12ms	253 MB
8	8,745	141.43 GB	54.29ms	320 MB
16	16,734	270.05 GB	56.47ms	453 MB
32	25,555	412.36 GB	72.48ms	699 MB

Near-linear scaling up to 16 threads in both modes.

Differential Revision: D106447554

[meta-codesync]

@xiaoxmeng has exported this pull request. If you are a Meta employee, you can view the originating Diff in D106447554.