perf(parquet): batched DELTA decoding with SIMD bit-unpack

[jaylisde]

Summary

Closes the DELTA-vs-PLAIN/DICT scan gap on TPC-H Q12 (DELTA-encoded lineitem) left after facebookincubator#17633. DELTA scan CPU was still ~4.5x of PLAIN/DICT; after this PR it's ~1.8x.

The work covers both DELTA-encoded readers in Velox: DeltaBpDecoder (used by DELTA_BINARY_PACKED integer columns and as a length decoder inside DELTA_LENGTH_BYTE_ARRAY / DELTA_BYTE_ARRAY) and DeltaByteArrayDecoder (string columns). Three layered changes:

1. SIMD bit-unpack kernels for whole miniblocks at bit_width 0–32.
2. A batched decode fast path that calls ColumnVisitor::processRun() once per chunk instead of per-row process(), with SIMD filter pushdown for deterministic integer filters.
3. DeltaByteArrayDecoder micro-optimizations: devirtualization, decoder reuse across pages, and std::string → raw buffer.

PLAIN/DICT path is unchanged within noise.

Performance

TPC-H Q12 SF10, num_drivers=4 --num_repeats=5 --warmup_after_clear=true, 5-run median:

	Wall	lineitem scan CPU
baseline (facebookincubator#17633)	1.02 s	2.76 s
This PR	0.594 s	1.11 s
Δ	−42%	−60%

Implementation

1. SIMD Bit-Unpack Kernels

Compile-time dispatch on bit_width 0–32:

bit_width	Strategy
0	Arithmetic-sequence fast path (all deltas identical)
1–16	4 values/iter via single unaligned 64-bit load + shift/mask
17–32	2 values/iter via __uint128_t funnel-shift (lowers to SHRD on x86_64)
33–64	Scalar inner loop (no SIMD batching: a single value would span > 64 bits after byte-misalignment, rare in practice)

2. Batched Decode Fast Path (processRun)

Instead of per-row visitor.process(), decode an entire miniblock into the output buffer and call visitor.processRun() once per chunk. Two fast paths inside:

• AlwaysTrue (no filter): zero-copy, just advance the counter
• Deterministic filter + integer types: SIMD batch comparison via processFixedFilter

3. Sparse Buffered Path

For filtered reads, decode kBatch=1024 physical values into a stack buffer (≤ 8 KB for int64_t, well within typical thread stack limits), then index into it by rows[k] - batchPhysStart. One sequential decode pass (cache-friendly) turns random row access into simple array indexing.

4. DeltaByteArrayDecoder Optimizations

• Devirtualization: remove DeltaByteArrayDecoderBase, eliminate vtable dispatch from per-row readString()
• Allocation reuse: child RleBpDecoder uses std::optional + reset() instead of make_unique per page (saves 3 heap allocations per page boundary)
• Buffer optimization: replace std::string lastValue_ with std::vector<char> + length — eliminates std::string::_M_replace overhead (~6.5% in profiling) from prefix-share encoding