Clamp negative SDE sampler tmp

[fallintoplace]

Summary

Clamp the SDE sampler's intermediate tmp term before sqrt() so tiny negative drift cannot turn the update into NaNs.

Root cause

In the SDE branch, the sampler computes tmp = 1 - sigma_t**2 / (sigma_prev**2 + eps) and passes it directly to sqrt(). The shipped schedules stay non-negative in the normal path, but any slight monotonicity drift from finite precision or edge-case overrides can push tmp just below zero and poison the sample with NaNs.

What changed

• clamp tmp with clamp_min(0) before using it in the SDE update
• add a focused regression test that constructs adjacent timesteps near the boundary and forces a tiny negative pre-clamp tmp
• verify the sampler output stays finite and matches the clamped update behavior

Impact

The SDE sampler is more numerically robust around boundary steps and invalid-nearby schedules, without changing the normal monotonic cases.

Validation

• uv run pytest tests/collections/audio/test_audio_parts_submodules_schroedinger_bridge.py -k "clamps_negative_tmp_before_sqrt" -v
• uv run pytest tests/collections/audio/test_audio_parts_submodules_schroedinger_bridge.py -v
• uv run --with black==24.10.0 black --check nemo/collections/audio/parts/submodules/schroedinger_bridge.py tests/collections/audio/test_audio_parts_submodules_schroedinger_bridge.py
• uv run isort --check-only nemo/collections/audio/parts/submodules/schroedinger_bridge.py tests/collections/audio/test_audio_parts_submodules_schroedinger_bridge.py

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.