[LLM:Bugfix] Fix QNN export chunk-size and per-task workdir handling

[LudovicoYIN]

Description

This PR contains two small bug fixes in the LLM QNN export path.

1. Honor --chunk_size when generating prefill IO

transformers/llm/export/npu/generate_llm_qnn.py exposed --chunk_size, but convert_llm() always called:

makeIOJson(args, 128, hidden_size, mask_type)

This meant the prefill QNN graph could still be generated against 128-token IO even when --chunk_size 256/512 was requested.

2. Run QNN compile subprocesses inside the per-task workdir

source/backend/qnn/npu_convert.py invoked the compile subprocess without setting cwd=workdir.
For parallel conversions, this could make multiple tasks share the same process working directory assumptions and increase the chance of temporary/output path collisions.

Why this matters

In the MNN QNN offline export path, these issues can make generated artifacts inconsistent with runtime expectations:

• prefill graph shape may not match the requested chunk size
• runtime chunk_limits and generated graphs may diverge
• ppl_eval results can become misleading
• parallel export jobs are more fragile than they need to be

Module

LLM

Type

• Feature
• Bugfix
• Perf
• Refact
• Style
• Doc
• Test
• Chore

Checklist

• Commit message follows [Module:Type] Description format
• Code compiles without errors
• Tested on relevant platform(s)
• No unrelated format or style changes included

Fixed

#4612

[LudovicoYIN]

thanks your review.