Clarify positional encoding length limit errors - #3085

What does this PR do?

Related to #2071, specifically the requested diagnostic for inputs exceeding the positional encoding limit.

PositionalEncoding silently returns only max_len positions for a longer input. ASR callers then fail with a tensor-size mismatch that does not identify the configured limit. Raise ValueError in the shared module with both the input length and the maximum length, before returning an incomplete encoding.

The configured capacity and sinusoidal values stay unchanged. Inputs at the limit remain valid. This does not add support for transcribing arbitrarily long recordings.

Validation on CPU, Python 3.11.15 / PyTorch 2.6.0:

• New tests cover empty/short/exact-limit sequences, float32/float64 module conversion, custom/default limits, and real small TransformerASR encode/source/target paths. Before the fix: 5 failed, 8 passed. Afterward: 13 passed.
• New and related attention, conformer, streaming and mask tests, plus all Transformer doctests: 59 passed.
• Separate baseline comparisons give identical normal ASR outputs, input gradients and checkpoint metadata. TorchScript retains valid-boundary behavior and reports the new error for over-limit inputs.
• Full-repository and changed-file pre-commit checks and git diff --check pass.

The tests use synthetic features and a locally initialized small model. No pretrained models, datasets, GPU execution or full repository test suite were used. Existing mask-type deprecation and attention-test overflow warnings remain.

Implemented and tested autonomously with OpenAI Codex, with independent read-only review by another Codex agent.

PR review