`Mask::valid_counts_for_indices` walked the validity bit buffer one bit at
a time via `BitIterator::next`, accumulating a running valid count. The
pco/zstd slice decoders call it with `[slice_start, slice_stop]`, so the
cost was a bit-by-bit scan of the entire prefix up to `slice_stop`.

Replace the per-bit walk with an incremental SIMD popcount over each gap
between consecutive indices. Total scanned bits are unchanged, but the
work now goes through the vectorized `count_ones` path.

Add `BitBuffer::count_range(start, end)`, which counts set bits in a range
directly over the backing buffer without cloning a sliced `BitBuffer`, so
the many-indices case has no per-gap allocation overhead.

Benchmark (vortex-mask/benches/valid_counts.rs, median):

  slice_bounds 16384    8.7us  -> 51ns    (~170x)
  slice_bounds 262144   138us  -> 310ns   (~445x)
  slice_bounds 1M       554us  -> 1.86us  (~300x)
  many_indices 16384    10.1us -> 1.68us  (~6x)
  many_indices 1M       613us  -> 3.33us  (~185x)

Signed-off-by: Joe Isaacs <joe.isaacs@live.co.uk>
Signed-off-by: Robert Kruszewski <github@robertk.io>