Strict Sphinx build has duplicate objects and unresolved PyTorch references

### Description

After the malformed LittleKernel table is fixed separately in #188 / #189, the strict documentation build at current main commit `8260bc34398c2b8f36dc840fd22f741ca9294584` still fails with 26 warnings:

```console
sphinx-build -E -a -n -W --keep-going -b html docs _build/html
build finished with problems, 26 warnings (with warnings treated as errors).
```

The warnings are:

- 5 duplicate object descriptions for `EPAllToAllLayoutDesc` and its attributes, defined in both `layers/nvidia/ep_a2a_layer.rst` and `layers/nvidia/ep_a2a_fused_layer.rst`
- 21 unresolved `py:class` references to `torch.Tensor` or `torch.dtype` across `low_latency_a2a_v2.rst`, `ep_a2a_fused_layer.rst`, `ep_a2a_layer.rst`, and `ep_ll_a2a_layer.rst`

### Expected behavior

A strict full Sphinx build should complete without warnings. For the duplicate objects, the docs should either define the type once or deliberately mark one copy as non-indexed. For PyTorch types, the project should either configure an authoritative inventory (for example, intersphinx) or render the types without unresolved cross-reference roles.

I am leaving this issue patch-free because both warning groups require a documentation-policy choice rather than a mechanical suppression.

Run the strict command `sphinx-build -E -a -n -W --keep-going -b html docs _build/html` on the stated commit and review warnings in `low_latency_a2a_v2.rst`, `ep_a2a_fused_layer.rst`, `ep_a2a_layer.rst`, and `ep_ll_a2a_layer.rst`. Decide how the duplicate `EPAllToAllLayoutDesc` definitions and unresolved `torch.Tensor`/`torch.dtype` references should follow the documentation policy. Done means the full strict build completes without warnings.

Written by the indexing model from the issue text.