[Tianlei Wu (tianleiwu)]

Summary

Extends com.microsoft::SpeculativePrefixCommit (introduced in #33474) to support batch_size >= 1 for batched speculative decoding:

• Schema and shape inference (bert_defs.cc): Propagates the batch dimension from next_tokens ([batch, rows] or [rows]) to outputs 0..6 (head_ids, head_c, slot_ids, seqlens, cos_out, sin_out, sel).
• CUDA kernel (speculative_fold.h/.cc/.cu): Launches MidKernel<<<batch_size, 256, 0, stream>>> with per-slot strides for ctl, ctx_commit, prefixes, prefix2, per-slot result ring buffers, and per-slot muted flags.
• Unit test (speculative_fold_op_test.cc): Adds SupportsBatchedSpeculativePrefixCommitS8 covering batch = 3, head_rows = 8 with mixed acceptance lengths and a muted slot.