perf: speed up distinct index collection in low-memory pipeline

[hjanuschka]

This removes Option staging in get_distinct_indices and builds the result directly in position order. It reduces overhead in a hot helper used during low-memory pipeline row setup without changing behavior. Unsafe is used only to reborrow previously collected mutable pointers, and is safe because indices are distinct and sorted, so aliasing cannot occur.

[github-actions]

Benchmark @ e39e2c5

MULTI-FILE BENCHMARK RESULTS (8 files)
  CPU architecture: x86_64
  WARNING: System appears noisy: high system load (2.03). Results may be unreliable.
Statistics:
  Confidence:               99.0%
  Max relative error:        3.0%

Comparing: 462454f5 (Base) vs cfa1595f (PR)

File	Base (MP/s)	PR (MP/s)	Δ%
bicycles.jxl	6.975	6.982	+0.10% ±0.7%
bike.jxl	24.708	24.938	+0.93% ±1.7%
delta_palette.jxl	6.101	6.128	+0.43% ±0.3%
green_queen_modular_e3.jxl	8.259	8.194	-0.79% ±1.1%
green_queen_vardct_e3.jxl	25.305	25.187	-0.46% ±1.9%
lz77_flower.jxl	3.316	3.353	+1.13% ±0.3%
patches_lossless.jxl	3.178	3.176	-0.05% ±0.8%
sunset_logo.jxl	2.665	2.744	+2.95% ±1.1%

[veluca93]

Please (1) add an arbtest that checks that the code is fine (2) run this test under miri

[veluca93]

This also needs a benchmark, enqueued it :-)