[Bug][AutoDeploy]: Investigate _scratch_block_offsets in trtllm_attention.py

Problem

We observed crashes without a defensive copy of block_offsets into a scratch buffer before each thop.attention call. This occurred with Llama + TRTLLM Attention + overlap scheduler ON + torch-cudagraph. Adding the workaround described below fixed it - we should investigate why this was.

The hypothesis is that metadata planning for an upcoming batch can write to shared buffers (like block_offsets) while a concurrent thop.attention call is still reading from them.

Current workaround

A _scratch_block_offsets buffer is allocated in _GlobalTrtllmPlanner.init_spec_decoding() and used in trtllm_mha_with_cache (gated on is_spec_decoding_enabled). See trtllm_attention.py.

TODO

• Understand the root cause of the race more deeply
• Determine if non-spec-dec paths (vanilla Llama + overlap scheduler) also need the scratch copy
• If so, move the copy out of the spec-dec gate — this may fix the known torch-simple + overlap scheduler crash. If not, diagnose the root cause and fix it instead of this scratch block offsets.