add support for mlp and kalman to sample_adapt_regressor

[kylmcgr]

Sample adaptive regressor works in the cursor pipeline with Kalman and MLP (torch) models.

[cboulay]

1. I don't really like having multiple inert units and then only wiring up one of them. This can lead to mistakes. e.g., you rewired the signal but you did not rewire the sampler: (self.SEQSEQSAMPLER.OUTPUT_SAMPLE, self.REGRESSOR.INPUT_SAMPLE) has no equivalent for the other backends.
   I have another pattern that you can use that I will send you.

2. I'm not sure about the shape and axes. If we use a non-zero decode_window_dur then the windower's output has a leading "win" axis. Then we flatten, and honestly I can't remember what that does but it looks like preserve_axis="win" would keep the win axis? So when we get to DecodeOputputAdapterProcessor, it only brings the time axis through if there is a "time" axis. Do we have that?