Remove non-finite SSFT noise in StochasticNoise

[mo-AliceLake]

Addresses https://github.com/metoppv/mo-blue-team/issues/1220

Stochastic noise generation can occasionally produce non-finite values (NaN/Inf), which then propagate into output. This leads to sporadic failures in stochastic-noise tests (see e.g. example 1, example 2).

To fix this, updated improver/calibration/stochastic_noise.py to guard against non-finite values:

• In _from_dB: treat non-finite linear values as below-threshold by setting them to 0.0 before threshold masking.
• In _process_single_realization: replace any remaining non-finite values in noise_linear with 0.0 before scaling etc.

Updated improver_tests/calibration/stochastic_noise/test_StochasticNoise.py:

• Added test_process_non_finite_noise_is_sanitized
• Test forces do_fft output to include NaN/Inf and verifies:
  • Non-positive-region output is finite
  • Positive-region values remain unchanged
  • With scaling enabled, non-positive-region values are <= 0

Testing:

• Ran tests and they passed OK
• Added new tests for the new feature(s)

[MoseleyS]

Looks good

[brhooper]

Thanks @mo-AliceLake

This change should solve our issues with non-finite values being generated and the new functional is covered by a new unit test. I've added one comment to fix a typo, then I think that this can be merged.

[brhooper]

Looks good to me. Thanks @mo-AliceLake .