Fix temporal BEV alignment for batched tracking

[zjn20030811]

Fixes #268

Summary

• Select each sample's previous BEV from the interleaved [previous, current] temporal queue before constructing the temporal-attention query.
• Preserve the queue layout used by the deformable-attention value path while avoiding cross-sample history leakage for batches larger than one.
• Add deterministic numerical regression coverage for sample alignment, the attention forward path, and malformed queue lengths.

Validation

• python -m pytest -q tests/test_temporal_self_attention.py (3 passed)
• python -m py_compile projects/mmdet3d_plugin/uniad/modules/temporal_self_attention.py tests/test_temporal_self_attention.py
• ruff check tests/test_temporal_self_attention.py
• git diff --check

The attention regression uses a deterministic deformable-attention substitute only at the optional extension boundary; the queue selection and query construction execute the production implementation. With the pre-fix slice, sample 1 receives sample 0's current BEV; the test now verifies [previous_0, previous_1] alignment.