Vulkan:Speed: Speed up vulkan for corp-matrix, complete segment ops

[jxt1234]

segment Vulkan buffer 2194.10 tok/s 86.32 tok/s
non-segment Vulkan buffer 1100 tok/s 55 tok/s

Description

Module

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

[wangzhaode]

Thanks for the substantial Vulkan buffer optimization. The reported gains (prefill 1100 → 2194.10 tok/s, decode 55 → 86.32 tok/s) look promising.

Could you please make a few updates before we merge?

1. Please rebase onto the latest master and resolve the current conflicts.
2. Please keep the shader-generation changes focused. The current diff changes subprocess.run back to os.popen in three makeshader.py files and also includes changes under skills/. Please revert these unless they are required for this optimization, or briefly explain why they are needed.
3. Please add the basic benchmark context (device/GPU, model, quantization, and build flags) together with a simple CPU vs Vulkan correctness result.
4. Please confirm that changing the MNN_VULKAN_IMAGE default to OFF and unconditionally enabling MNN_GPU_RECORD_BATCH are intentional. The embedded SPIR-V size grows from about 1.41 MB to 1.94 MB (about +527 KB), so a brief binary-size note would also help.

No exhaustive device matrix is required; one representative real-device correctness/performance result is enough. Thanks!