[ColinLeeo]

What changes

RLE was the last hot value codec without a batch path: Plain, Gorilla and
TS_2DIFF received read_batch_* in #823/#898, while RLE still decoded value
by value through read_int and unpacked each bit-packing group with the
portable scalar packer. On a wide table (thousands of measurement columns per
device) the RLE decoder dominates the read profile.

• Batch decode — read_batch_int32 / read_batch_int64 and
  read_batch_float / read_batch_double. An RLE run becomes a std::fill, a
  bit-packing group becomes one memcpy out of the unpacked buffer.
• SIMD bit-unpack — four values per step with a gather+shuffle kernel, plus
  a native AVX2 path dispatched at runtime via __builtin_cpu_supports instead
  of relying on the translation unit being compiled with -mavx2. The portable
  simde fallback stays for other targets. Widths that can cross a ninth byte
  (59/61/62/63) use a 128-bit buffer where the compiler supports __int128.
• Less copying — each bit-packing group is read with one bulk read instead
  of per-byte read_ui8, and the pattern length wraps the caller's ByteStream
  instead of copying it into a temporary buffer.

Correctness notes

• The SIMD gather reads up to 7 bytes past the last packed group, so buffers
  must hold 8 readable bytes of padding after the final group. The RLE decoders
  already zero-pad; the Sprintz block decoders now do too, and the packer tests
  pad their buffers and pass the exact packed length.
• MSVC has no __int128, so the 128-bit scalar path is guarded by
  __SIZEOF_INT128__ with the portable bit-by-bit loop as the fallback.

Measurements

Unpack micro-benchmark, values/s, -O2 without -mavx2 (runtime dispatch
takes the AVX2 path on the test machine):

bit width	before	after
14	~99 M	~547 M
20	~99 M	~519 M

End-to-end on a wide-table workload (5 features x 4500 measurement columns x
1 month of minute data): this step alone took a single query from 3.14 s to
2.38 s.

Testing

• Full C++ suite on this branch: 1428 tests pass.
• The same files also pass a 1440/1440 run under -Denable.asan=ON (Debug) as
  part of a larger branch, including LeakSanitizer.