perf: concatenate per-chunk Schur edge lists in one pre-sized copy

[schroedk]

par_emit collected sampled star edges with reduce(Vec::new, |a, b| a.append(b)), so rayon's reduction tree recopied every edge once per level. Collecting the chunks and calling concat() copies each edge exactly once into one pre-sized allocation.

Setup time on fixest_comparison, interleaved rounds with a rebuild between arms:

	500K	2M	5M	20M
LSMR(exact) (null control)	+1.5%	+2.1%	+1.4%	+0.3%
LSMR(approx)	-3.2%	-4.1%	-4.1%	-5.0%
LSMR(2-2)	-9.0%	-5.6%	-6.4%	-6.2%

Run ranges separate at 5M and 20M for approx, and at every size for 2-2.

Scope: null on slopes, akm_panel, akm_scaling and six small-problem suites — Schur assembly is ~30% of within CPU on fixest-difficult 3FE but only 5-7% there. No regression in any suite measured.

Results are unchanged: sort_and_dedup imposes a total (lo, hi, weight) order immediately after, and iteration counts are bit-identical across every case/config pair in all runs.

Note: 5% sits inside this repo's normal noise band, so reproducing it needs the interleaved protocol — a casual before/after will not show it.