[Train] Reduce padding-free embedding output memory

[taking-lying-flat]

What changed

• Select the last hidden state of every sequence directly from the packed embedding output.
• Keep the existing re-padding behavior for sequence classification and reranking tasks.
• Support both cu_seq_lens_q and packed position_ids, including the sequence-parallel path.
• Add regression tests for output selection and gradient propagation.

Why

With padding_free, the embedding path restored the full output to a dense
[num_sequences, max_sequence_length, hidden_size] tensor before the output
normalizer immediately selected only the last token of each sequence. For
multimodal batches with uneven sequence lengths, this creates a large temporary
allocation and retains its autograd graph until backward.

The embedding loss only needs one vector per sequence, so selecting those packed
positions directly avoids the re-padding allocation without changing the
resulting embeddings.

Fixes #9885.

Validation

• pytest -q tests/models/test_patcher.py
• pre-commit run --files swift/model/patcher.py swift/trainers/mixin.py tests/models/test_patcher.py
• Tiny Qwen3-VL packed forward, normalized output, and backward smoke test