[BUG] Propagate dtype conversion through composed and wrapped metrics

[AHMETHAKANBEZIR1]

LLM generated content, by OpenAI Codex

Reference Issues/PRs

Independent of #1897 (input dimensions) and #1754 (checkpoint device mapping).

What does this implement/fix? Explain your changes.

MultiLoss([MAE(), MAE()]).set_dtype(torch.float64) and (MAE() + MAE()).set_dtype(torch.float64) currently change the composite's dtype but leave constituent metrics and state in float32. These metrics are held in ordinary Python lists. For a third-party TorchMetrics metric accepted by the public coercion factory, converting its wrapper also leaves the underlying metric's state in the old dtype because TorchMetrics requires its own set_dtype call.

Delegate set_dtype through MultiLoss, CompositeMetric and TorchMetricWrapper, retaining the parent conversion and fluent return value. This preserves public list representation, registration/state-dict behavior, weights and each metric's conversion semantics. Device conversion and checkpoint loading are unchanged.

What should a reviewer concentrate their feedback on?

Explicit delegation through each composition boundary, including conversion after update/compute and defaults used by reset. No module-registration redesign is involved.

Did you add any tests for the change?

Twenty-five cases cover float32/float64 in both directions, no-op conversions, empty/populated state, cached compute results, reset defaults, third-party TorchMetrics wrapped through the public factory, and the string dtype API with adapted MSELoss, weighted results and analytic gradients. Clean main 6d0d18a has 13 failures and 12 passing controls; all twenty-five pass with the fix.

Complete tests/test_metrics.py: 95 passed, 6 existing optional cpflows skips. Configured pre-commit hooks pass across all files and at commit. CPU Python 3.12.14 / PyTorch 2.13.0. Full-package, GPU, distributed and training benchmarks were not run.

Any other comments?

Prepared autonomously with OpenAI Codex; no independent human review is claimed.

PR checklist

• The PR title starts with [BUG].
• Added/modified tests.
• Used pre-commit hooks when committing.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@6d0d18a). Learn more about missing BASE report.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2467   +/-   ##
=======================================
  Coverage        ?   89.07%           
=======================================
  Files           ?      218           
  Lines           ?    11684           
  Branches        ?        0           
=======================================
  Hits            ?    10407           
  Misses          ?     1277           
  Partials        ?        0           

Flag	Coverage Δ
cpu	89.07% <100.00%> (?)
pytest	89.07% <100.00%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.