Support flexible EpochMetric compute_fn outputs

[Abhisri436]

Fixes #1757

Description:

EpochMetric is currently typed and documented around scalar outputs, while issue #1757 requests support for richer compute_fn return types. My investigation also showed that unsupported values such as strings could pass through without clear validation in single-process mode.

This PR extends EpochMetric.compute_fn output handling beyond scalar values. It adds validation for supported output types including int, float, torch.Tensor, sequences, and mappings with string keys. Unsupported output types now raise a clear TypeError.

It also adds tests for scalar tensor, vector tensor, tuple/list tensor outputs, mapping outputs, and unsupported output behavior.

For distributed mode, this implementation takes a conservative approach: scalar and tensor outputs can still be broadcast, while tuple/list/mapping outputs raise a clear NotImplementedError when world_size > 1 instead of attempting unverified recursive distributed container broadcasting.

Validation:

• Before: unsupported output types such as str could return without a clear error in single-process mode.
• After: supported tensor, sequence, and mapping outputs are validated and tested, while unsupported output types raise a clear TypeError.
• pytest tests/ignite/metrics/test_epoch_metric.py -vv → 17 passed, 57 skipped
• WORLD_SIZE=2 distributed tests with the gloo backend → 10 passed
• The distributed tests verify that non-scalar tensor outputs are consistent across ranks and that unsupported container/type outputs fail symmetrically without mismatched collective calls.
• ruff check ignite/metrics/epoch_metric.py tests/ignite/metrics/test_epoch_metric.py → All checks passed
• ruff format --check ignite/metrics/epoch_metric.py tests/ignite/metrics/test_epoch_metric.py → 2 files already formatted

Note: I saw that PR #3789 also addresses this issue. I had already selected and started working on issue #1757 before discovering the overlapping PR. I am submitting my completed branch for review and am happy to adjust this PR if maintainers prefer.

[Abhisri436]

Hi maintainers and reviewers! I’ve added permanent WORLD_SIZE=2 distributed tests and updated the PR with the latest validation results.

I also compared the distributed behavior with the overlapping PR and found a meaningful difference in how multi-rank outputs are handled. I’ve documented the tested behavior and added regression coverage in commit fcaef4cf.

I would appreciate a review when someone has time, and I’m happy to adjust or consolidate the approach if preferred. Thank you!