Fix inf rolling averages passed to compute_softmax

[sniper-noob]

Problem:
When a miner has no valid scores in the moving average window, compute_smoothed_score assigns rolling_avg = float("inf") (line 178). These inf values are passed directly to compute_softmax (line 199).

With a single inf, softmax works by accident: beta * inf = -inf, exp(-inf) = 0, so the inf miner gets 0 weight. But when ALL miners have inf (e.g. all are new with no scores yet), softmax breaks:

scaled = -0.2 * [inf, inf, inf] = [-inf, -inf, -inf]
shifted = [-inf] - max([-inf]) = [nan, nan, nan] (inf - inf = nan)
exp([nan, nan, nan]) = [nan, nan, nan]
weights = [nan, nan, nan] / nan = [nan, nan, nan]

NaN weights crash downstream reward distribution.

Even with a single inf, the behavior is fragile: the inf miner gets exactly 0 weight instead of the worst-but-nonzero weight that a miner with the worst real score would get. This silently excludes them from rewards rather than giving them the minimum.

Fix:
moving_average.py lines 184-193:
Before passing to compute_softmax, replace inf rolling averages
with the worst (highest) finite rolling average. This gives
no-score miners the same weight as the worst real performer,
and ensures compute_softmax always receives finite inputs.