fix: reduce flamelet test to 1 step to avoid Cray GPU SIMD FP divergence

Cray ftn GPU-OMP uses 'target teams distribute parallel do simd' which
applies SIMD vectorization with GPU-specific transcendental math. Over
50 steps of stiff H2/O2 chemistry, tiny FP differences in GPU exp/log
accumulate to rel errors of ~1e-3 to ~8e-2, far exceeding any useful
tolerance. AMD GPU and all CPU builds agree with the gfortran golden.

Reducing to 1 step keeps FP accumulation negligible and still tests IC
loading, CBC boundary conditions, and chemistry kernel invocation.

Author: sbryngelson