[Issue]: Workspace Buffer Pointer Arithmetic in pa_v1 Ignores MTP

Problem Description

The pa_v1 C++ kernel computes workspace pointers without considering MTP
https://github.com/ROCm/aiter/blob/d2f5f2791336a9839c65241b77a88f0a8407a054/csrc/cpp_itfs/pa/pa_v1.cpp.jinja#L69-L72

This allocates based on num_seqs only. However, the main kernel writes to index seq_idx + mtp * MTP_PARALLEL_THREADS,
which for mtp > 0 can write beyond the allocated buffer, causing memory corruption.

While the Python test code correctly sizes the buffer with mtp:
https://github.com/ROCm/aiter/blob/d2f5f2791336a9839c65241b77a88f0a8407a054/op_tests/test_pa_v1.py#L308-L319

Also note that there is zero test coverage for MTP > 1:
The test file (test_pa_v1.py) never tests with mtp > 1

Operating System

Ubuntu 22.04.5 LTS (Jammy Jellyfish)

CPU

AMD EPYC 9575F 64-Core Processor

GPU

8 x AMD Instinct MI355X

ROCm Version

ROCm version: 7.0.51831-a3e329ad8

ROCm Component

No response

Steps to Reproduce

Replicator script:

paged_attention_v1.issue3.mi355.py

Logs from running replicator script:

paged_attention_v1.issue3.mi355.log

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

My goal is to enable CUDA graph support for speculative decoding in the ROCM_AITER_FA attention backend by adding spec_as_decode support and changing _cudagraph_support from UNIFORM_SINGLE_TOKEN_DECODE to UNIFORM_BATCH. This is expected to speed up speculative decoding: removes graph breaks and allows routing speculative tokens through the more efficient decode path instead of the extend path. The AITER PA API appears to support MTP>1, but the implementation seems to be buggy.

Start with csrc/cpp_itfs/pa/pa_v1.cpp.jinja lines 69-72 and compare its workspace sizing with the indexing described in the issue. Review op_tests/test_pa_v1.py lines 308-319, add coverage for mtp > 1, and run the pa_v1 tests to verify the workspace remains large enough without memory corruption.