Resolve Speculative Decode RTE

[tannervoas742]

• Speculative decoding fails when batch size exceeds 1 due to incorrect handling of mixed speculative and non-speculative sequences in the batch.
• This PR corrects batch expansion ordering and accounts for padding sequences.

Below is a reference to the line where we combine speculative and non-speculative. The order is clearly non-speculative first followed by speculative.

vllm-fork/vllm/spec_decode/batch_expansion.py

Line 135 in 4d91f3b

target_seq_group_metadata_list = non_spec_seqs + spec_expanded_seqs

Below is a reference to the line where we pad the batch with dummy sequences. These also must be accounted for.

vllm-fork/vllm/worker/hpu_model_runner.py

Lines 1270 to 1275 in 4d91f3b

	batch_size_padded = self.bucketing_ctx.get_padded_batch_size(
	real_batch_size, False)
	batch_size_padding = batch_size_padded - real_batch_size
	if batch_size_padding > 0:
	encoder_seq_lens.extend(encoder_seq_lens[0]
	for _ in range(batch_size_padding))

Below is the error that is encountered without this fix.

With this fix significantly higher throughputs can be achieved and accuracy is unimpacted (examined bert F1 accuracy; tested llama-3.1-8B with n-gram speculative decoding).