Hardware acceleration support and NSampler parallelization

[ben-nowacki]

Description

NSampler (and Paired/Triplet subclasses) now support parallelization via an n_workers argument in __init__.

• Anchor samples are batched into n_workers, where each worker independently find valid pairs for its own anchor samples.
• Worker batches are then joined on completion and re-sorted by anchor indices.
• Fixed phases execution clearing sampler outputs on every call, even if the sample configuration did not change. Sampler outputs are now effectively cached, and any sampler computation is only re-run if the phase input bindings or sampler configuration changes.

Hardware acceleration:

• New Accelerator class provides unified implementation of hardware acceleration for PyTorch and Tensorflow models.
• Supports CPU (default), GPU (and CUDA), and MPS (Apple Silicon; PyTorch only)
• All BatchViews produced by a phase's sampler are now pre-materialized before any training/evaluation occurs. This adds a slightly overhead at the start of each phase but removes redundant re-materialization on every epoch of training.
• All allocation of data and models onto defined devices (eg, GPU) now occurs during this materialization stage prior to phase execution
• A new how-to notebook has been added to the documentation outlining Accelerator usage

Related Issues

Closes #21
Closes #22

How Has This Been Tested?

• Local pytest run (No new unit tests added but new example notebook executes).
• CI passes (all nox tests pass)
• Manual checks
  • GPU and MPS support tested locally with PyTorch ModelGraph.
  • NSampler with 4 workers reduced sampling time from ~50 seconds to <10 seconds.
  • Execution of same TrainingPhase and ModelGraph setup reduced from ~3 minutes to <30 seconds with Accelerator("gpu:0", pin_memory=True)

Checklist

• My code follows project style guidelines: nox -s pre-commit
• I have added tests that prove my fix is effective or my feature works: nox -s unit
• I have updated documentation if needed: nox -s docs
• I have linked related issues

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 13.52740% with 505 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
modularml/samplers/n_sampler.py	4.34%	132 Missing ⚠️
modularml/core/topology/compute_node.py	13.46%	90 Missing ⚠️
modularml/utils/nn/accelerator.py	28.44%	78 Missing ⚠️
modularml/core/topology/model_node.py	8.43%	76 Missing ⚠️
modularml/core/experiment/phases/train_phase.py	7.14%	39 Missing ⚠️
modularml/core/topology/model_graph.py	16.66%	30 Missing ⚠️
modularml/core/experiment/phases/eval_phase.py	6.89%	27 Missing ⚠️
modularml/core/experiment/phases/fit_phase.py	13.04%	20 Missing ⚠️
modularml/core/sampling/base_sampler.py	14.28%	6 Missing ⚠️
modularml/core/experiment/phases/phase.py	20.00%	4 Missing ⚠️
... and 2 more

📢 Thoughts on this report? Let us know!