fix: discard cached gram matrix when refitting CostRbf/CostCosine

[gaoflow]

Closes #372.

CostRbf and CostCosine compute their Gram matrix lazily and cache it in self._gram, but fit() never invalidates the cache. Detection estimators create the cost object once in __init__ and call cost.fit(signal) on every fit, so refitting the same estimator on a different signal silently reuses the previous signal's Gram matrix:

m = rpt.Pelt(model="rbf")
m.fit(s1).predict(pen=3)  # [50, 100]
m.fit(s2).predict(pen=3)  # [50, 100]  <- stale; a fresh instance gives [30, 100]

For CostRbf there is a second layer: when gamma=None, the median-heuristic gamma is derived inside the gram property and stays pinned to the first signal, so resetting the cache alone is not enough.

fit() now discards _gram and re-derives the heuristic gamma per signal, keeping a user-supplied gamma — the same pattern CostMl already uses with has_custom_metric.

Tests: a parametrized refit-equals-fresh-fit check over all cost models, a gamma re-derivation/persistence test, and an estimator-level check for rbf/cosine (5 of these fail without the fix). Note: #369 touches costrbf.py for memory optimization but does not address this — its fit also keeps the cached state.