fix(strategies): refuse to tighten dynamic bounds on zero-range batch

[GrigoryEvko]

A zero-range batch (single sample, or all-equal samples) collapses the dimension. _calculate_margin(0) falls back to 1e-5, so for one elite with speedup=0.964 against initial bounds [0, 100], the bins become [0.96399, 0.96401] — a 2×10⁻⁵ window across 10 bins, five orders of magnitude tighter. Subsequent optimize_space calls with still-narrow batches re-collapse on every pass. Matches the [0.000, 100.000] -> [0.964, 0.964] log line in single-elite runs.

Reproducer against current main:

from gigaevo.evolution.strategies.models import DynamicBehaviorSpace, LinearBinning

space = DynamicBehaviorSpace(
    bins={"x": LinearBinning(min_val=0.0, max_val=100.0, num_bins=10)},
    expansion_buffer_ratio=0.1,
)
print(space.calculate_optimized_bounds([{"x": 0.964}]))
# before fix: {'x': (0.96399, 0.96401)}
# after fix:  {'x': (0.0, 100.0)}

Fix: when value_range == 0, return the strategy's current bounds unchanged. Two pre-existing tests (test_batch_optimize_single_point, test_single_point_batch) encoded the collapsed bounds as expected output and are updated.