Move NoiseGeneratorRecording from core to generation

[h-mayorquin]

Addresses the design discussion in #4482.

NoiseGeneratorRecording mixed two purposes: a lightweight lazy recording for testing and a simulation tool with spatial correlations (cov_matrix) and per-channel noise_levels. This PR introduces a minimal MockRecording in core/generate.py (unit-variance white noise only, no extra parameters) and moves the full NoiseGeneratorRecording to generation/noise_tools.py, where generate_noise() already lived. generate_recording() and generate_recording_by_size() now use MockRecording, while generate_ground_truth_recording() lazy-imports NoiseGeneratorRecording from generation. That way, the work on #4482 can proceed from here without interfering with the testing functionality. I will migrate the tests in a follow-up PR so this diff stays as clear as possible with minimal code churn.

Backward compatibility is preserved via module-level __getattr__ in both core/__init__.py and core/generate.py. Importing NoiseGeneratorRecording from the old paths still works but emits a FutureWarning directing users to import from spikeinterface.generation instead, with removal targeted for 0.106.0. I also changed SilencedPeriodsRecording to compose MockRecording + ScaleRecording instead of importing NoiseGeneratorRecording from core, to comply with a design decision once discussed with @samuelgarcia about the way modules should depend on one another (preprocessing should not depend on generation). I am happy to import from generation directly if you think that is better.

MockRecording still exposes the strategy argument for now. I would like to remove it and keep only tile_pregenerated, but I opened #4522 to discuss whether the simulation side needs on_the_fly for non-repeating noise. I will clean this up in a follow-up once that discussion settles.