bench: multi-seed bench_dsa_select — kill single-input pivot-luck spike (#357)

[woolcoxm]

What

Fixes a measurement artifact in bench_dsa_select that @KingIcyCreamProjects caught on the 9950X3D (see #357 thread): the nk=8192 cell reported ~75×, far above the real ~13–40× algorithm. Two compounding bench bugs, both verified against the code:

1. One frozen input per cell. partial_select_desc's median-of-three pivot is fully deterministic (no RNG), and the bench froze the input then ran 2000 reps on that identical array — so all reps hit the exact same pivot sequence. A single lucky input spiked one nk row (stable across re-runs because deterministic, not because real).
2. brng was never reset between cells. Each (shape, nk) cell's input depended on every prior cell's brand() draws — reordering nks[] or adding a shape silently shifted all later inputs.

Fix

• brng_seed() reseeds per (shape, nk, seed) → every cell reproducible and independent of cell ordering.
• Report the median of N_SEEDS=11 independent inputs, each itself a median over N_REPEAT=2000 timing reps. A lucky pivot now moves one of 11 samples, not the reported number.

Measured (this fix, Intel Core Ultra 9 185H, median of 11 seeds)

shape	nk	old	new	speedup
realistic	8192	631,600	57,600	10.97×
uniform	8192	833,700	88,800	9.39×
plateau	8192	161,100	10,000	16.11×

Band across all cells: 6–26×, no anomalous spikes. (plateau ignores the RNG so it's unchanged, as expected.) On the 9950X3D the 75× row should now land in the same band.

Scope

• Bench fidelity fix only. No engine code touched. partial_select_desc and attention_rows are unchanged.
• Correctness unaffected — test_dsa_select (129 cases) still passes.

Credit

The artifact and its diagnosis are @KingIcyCreamProjects's — caught in the #357 thread, with the correct root-cause read (deterministic pivot + single frozen input). This PR implements their suggested fix (average seeds per cell) and additionally fixes the never-reset-brng confound they didn't mention.