perf(postprocessing): compute only the amplitude scaling coefficients

[JESUSROYETH]

amplitude_scalings calls linregress per ordinary spike and creates a LinearRegression per collision, although only the fitted coefficients are used. This computes unused statistics and repeats estimator validation inside the per-spike loop.

This change computes the ordinary slope directly from centered covariance/variance in float64, and calls scipy.optimize.nnls directly for collisions. The float64 accumulation avoids the SciPy-version-dependent float32 rounding found during validation, while the collision path keeps the same intercept and non-negative constraint.

On a 384-channel, 200-unit, 120-second analyzer with 191,674 spikes (n_jobs=1), the public pipeline that feeds valid_unit_periods went from 110.224 s median (107.983–111.503) to 45.757 s (45.666–45.815), or 2.41x. The 191,674 scalings and collision flags were bit-identical. A public Spike2 recording with 23,096 events and 2,140 collision fits went from 9.386 s to 1.017 s (9.23x), also bit-identical. The focused tests pass, 9 of 9.