SparkDistill: 3x faster time-decay for merged-PR scores

[ai-hpc]

What

Add a scoring.time_decay override for gittensor-model-hub/SparkDistill so a merged PR's score multiplier decays 3x faster in time than the org default. Grace period, sigmoid midpoint, and sigmoid steepness are all scaled by 3x, which is an exact time-compression of the curve (g(t) == f(3t) against the previous curve) — the floor value is unchanged.

param	before (org default)	after
grace_period_hours	12	4
sigmoid_midpoint_days	10	3.33
sigmoid_steepness	0.4	1.2
min_multiplier (floor)	0.05	0.05 (unchanged)

Milestones (days since merge):

	before	after
50% of score remaining	day 10	day 3.33
~10% remaining	day 15.5	day 5.2
5% floor reached	day 17.4	day 5.8

Why

SparkDistill currently has no scoring.time_decay override, so a merged PR's score decays on the org-wide default curve (50% loss by day 10). That's slower than SparkDistill wants: verified distillation-quality improvements should reward fresh contributions more sharply, so miners are incentivized toward a steady stream of new verified improvements rather than living for a long time off one old merge.

Scope / safety

• Only gittensor-model-hub/SparkDistill's scoring.time_decay block is added — no other repo's config changes.
• emission_share, maintainer_cut, label_multipliers, and eligibility are untouched.
• All new values are within load_weights.py's live bounds (grace_period_hours in [0,168], sigmoid_midpoint_days in [1,90], sigmoid_steepness in [0.01,5], min_multiplier in [0,1]) — test_load_weights.py passes (132 passed).
• 9-line diff, JSON validates.