Add AFL fuzz target for prefix filter correctness

Oracle-based differential fuzzer: runs the same AFL-derived operation
sequence against two trees (one with prefix extractor, one without) and
asserts all reads return identical results. Any mismatch = wrongly
applied filter = silent data loss, saved by AFL as a crash for replay.

Covers all identified correctness dimensions:
- 9 extractor variants × 3 bpk levels × 3 filter partitioning policies
- MVCC snapshot reads at older seqnos while writes continue
- Weak tombstones and their compaction GC interaction
- Extractor changes on reopen (prefix_filter_allowed compatibility)
- Partitioned filter forced on all levels (the path that had the panic)
- Bidirectional iterator stepping (PrefixPingPong)
- Unbounded iteration (FirstKV/LastKV)
- Clustered keys (first byte 0..7, len 1..9) for realistic prefix
  distribution with natural in-domain / out-of-domain key mix

Author: zaidoon1

fuzz/prefix_filter/.gitignore

@@ -0,0 +1,2 @@
+in*
+out*

fuzz/prefix_filter/Cargo.toml

@@ -0,0 +1,10 @@
+[package]
+name = "prefix_filter"
+version = "0.1.0"
+edition = "2024"
+
+[dependencies]
+afl = "*"
+arbitrary = { version = "1", features = ["derive"] }
+lsm-tree = { path = "../.." }
+tempfile = "3.23.0"