Use channel offsets, not gains, when scaling amplitudes to uV

[JESUSROYETH]

SpikeAmplitudeNode and AmplitudeScalingNode read the channel gains into their offset array, so with return_in_uV=True (the default) they return raw * gain + gain instead of raw * gain + offset:

self._gains = recording.get_channel_gains()
self._offsets = recording.get_channel_gains()   # should be get_channel_offsets()

Every amplitude is therefore off by exactly gain - offset µV. On an Intan .rhd recording (gain=0.195, offset=-6389.76) that is +6389.955 µV on spikes whose true amplitude is about -46 µV, which also flips their sign; on a zero-offset recording such as SpikeGLX (gain=2.34375) it is a constant +2.34 µV on every spike.

Measured over 1443 spikes of a generated ground-truth recording, comparing each amplitude against recording.get_traces(return_in_uV=True) at the exact sample and channel that amplitude is taken from:

recording	error before	error after	median amplitude before → after (trace value)
Intan-like, gain 0.195, offset -6389.76	+6389.955 (constant)	2.6e-04	+6344.325 → -45.630 (-45.630)
per-channel gains and offsets	+5130.045 … +6933.117	5.9e-04	+5363.440 → -45.555 (-45.555)
SpikeGLX-like, gain 2.34375, offset 0	+2.34375 (constant)	0.0	-42.188 → -44.531 (-44.531)

amplitude_scalings passes the scaled waveform to linregress and keeps only the slope, so a uniform offset is absorbed by the intercept and the scalings do not move. When gains and offsets vary per channel the shift is not uniform and the median scaling goes from 2.5162 to 1.0018.

The scaling was correct before the node-pipeline refactor in d3ac13f, which replaced recording.get_traces(..., return_scaled=return_scaled) with the inline expression. Of the seven places outside tests that populate an offset, the other five all call get_channel_offsets().

One regression test is added to the existing test_spike_amplitudes.py, asserting that an amplitude equals the trace sample it is taken from. On unpatched main it fails for 141/141 spikes with a maximum absolute difference of 102.0, which is gain - offset for the test's gain 2.0 and offset -100.0.

Test suites on Linux / Python 3.12: postprocessing 98 passed 40 skipped, metrics + exporters 58 passed, curation + analyzer core 98 passed 1 skipped. Not run on macOS or Windows, and not on 3.10 / 3.13.