perf(parquet): SIMD bit-unpack for DeltaBpDecoder dense and sparse visitor paths

[jaylisde]

Summary

Builds on #3 (PR A) and #4 (PR B). Closes the remaining DELTA-vs-PLAIN/DICT scan gap on TPC-H Q12.

1. Inline SIMD bit-unpack in DeltaBpDecoder::decodeLongs. When the read aligns at a miniblock start and consumes a whole miniblock, dispatch on bit_width 0..32 to a compile-time-specialized kernel: bw=0 is an arithmetic-sequence fast path; bw 1..16 does 4-value/iter with a single unaligned 64-bit load (4×16 fits in one u64 window); bw 17..32 does 2-value/iter with __uint128_t funnel-shift (lowers to two u64 loads + SHRD on x86_64). The trailing u64 read is safe — intra-page overshoots fall into the next miniblock and the last miniblock has kPageReadPadding (8) trailing bytes guaranteed. bw 33..64 falls through to the existing scalar inner loop.

2. readWithVisitorSparseBuffered path for !Visitor::dense + !hasNulls + deterministic filter + NoHook + integral DataType (Q12's hot path, since chained filters produce a sparse row set after l_shipmode is applied). Decodes kBatch=1024 physical values into a stack buffer via decodeLongs (now SIMD), then walks the visitor's sparse rows[] array using rows[k] - batchPhysStart as buffer index. The existing per-row visitor.process is preserved — only the decode side is batched. n is capped to the visitor's residual physical span so the decoder never advances past what the visitor will consume.

3. readWithVisitorDenseBatched kBatch raised from 256 to 1024 to amortize the chunk-loop overhead now that decodeLongs is much faster per call.

Bench

TPC-H Q12 SF10, DELTA-encoded lineitem, num_drivers=4, 5-run median:

	Wall	lineitem TableScan CPU
#3 baseline (PR A)	0.85 s	2.13 s
+ #4 (PR B)	0.74 s	1.74 s
+ this PR	0.626 s	1.38 s

PLAIN/DICT parquet paths are unchanged within noise. Bench will be re-measured (5-run median) before posting upstream.

Test plan

• velox_dwio_parquet_delta_bp_decoder_test (new): 43 tests pass — 11 boundary cases (bit_widths 0/8/10/16/24/32, multi-block in a single readValues, mid-miniblock split across two calls, negative minDelta, bit_width > 32 fallback, narrowing to int32_t) plus 32 parameterized roundtrip tests covering bit_widths 1..32.
• velox_dwio_parquet_table_scan_test: 54 tests pass.
• velox_parquet_e2e_filter_test: 34 tests pass.

Part of #2.

[jaylisde]

Squashed into combined PR #6 (perf/delta-decoder-overhaul). Closing this stacked sub-PR — review continues on #6.