perf: packed hash encoding to reduce memory ~451 → ~240 B/key

[kacy]

summary

replaces HashValue::Compact(Vec<(CompactString, Bytes)>) with HashValue::Packed(Vec<u8>) using a listpack-style wire format. per-field overhead drops from 48 bytes to 6 bytes, cutting total per-key memory for the standard benchmark case (5 fields) from ~451 to ~240 bytes.

packed format: [num_fields: u16][{name_len: u16, name: bytes, val_len: u32, val: bytes}...]

• all buffer reads use checked accessors — truncated/corrupt buffers stop iteration gracefully
• get() returns Option<&[u8]> (zero-copy into the packed buffer); callers use Bytes::copy_from_slice at the keyspace boundary
• new field insert is append-only (no rebuild); existing field update or remove uses splice/drain on a small contiguous buffer
• HashMap promotion at >32 fields is unchanged
• no changes to command dispatch hot path, persistence wire format, or replication

what was tested

• cargo test -p emberkv-core — all 507 tests pass (including new packed-specific edge cases: truncated buffer, middle field removal, in-place update, roundtrip via to_hash_map/from)
• cargo clippy -p emberkv-core — clean, no warnings
• cargo build --workspace — full workspace compiles

design considerations

the packed buffer trades O(1) field update for O(n) scan + rebuild, which is acceptable because n ≤ 32 and the buffer is contiguous in L1 cache (~80 bytes for 5 fields). the read path (HGET) benefits from better cache locality — scanning 6 bytes of framing per field instead of 48-byte tuples. write_u16 uses direct indexing which is safe because it's only called on offsets that were just read successfully from the same buffer.