feat(tf2): support or reject legacy TensorFlow-only descriptors

## Summary

The unified schema exposes TensorFlow descriptor variants including `loc_frame`, `se_a_tpe`/`se_a_ebd`, `se_a_ebd_v2`/`se_a_tpe_v2`, and `se_a_mask`, while TF2 does not register matching descriptor implementations.

The current “TensorFlow” support label does not distinguish legacy TF from TF2, so users discover configurations that fail only after backend selection.

## Scope

For each descriptor family, either:

- implement a TF2-compatible descriptor with shared serialization/inference semantics; or
- add backend-aware validation and documentation that explicitly limits it to legacy TF.

Avoid silent substitution with a different descriptor.

## Acceptance criteria

- Every listed descriptor has an explicit TF2 support status.
- Supported descriptors construct, serialize, deserialize, and run deterministic forward/inference tests.
- Unsupported descriptors fail during backend-aware validation with a clear TF2-specific message.
- Aliases resolve consistently with the common schema.
- Documentation distinguishes legacy TF and TF2 support.

Refs #5757.

Coding agent: Codex
Codex version: codex-cli 0.144.4
Model: gpt-5.6-sol
Reasoning effort: xhigh