Feature/vulkan fa large buffer

[Yvi71]

Overview

This PR fixes a crash/abort in the Vulkan backend when utilizing large context sequence lengths (or high parallel slot counts) where the required intermediate split_k buffer or model inputs exceed the GPU's maxStorageBufferRange (capped at 4GB on AMD discrete GPUs like the RX 6800).
Although the scheduler's supports_op already correctly allows tensors larger than maxStorageBufferRange when shader_64b_indexing is enabled, the Flash Attention implementation (ggml_vk_flash_attn) was still hard-aborting on preallocation size limits and did not select the 64-bit indexing pipeline variants at runtime.
This change:

1. Relaxes the preallocation limit check to validate against max_buffer_size instead of maxStorageBufferRange if the device supports shader_64b_indexing.
2. Checks if 64-bit indexing is required based on the size of the bound buffers.
3. Retrieves and dispatches the 64-bit indexing variants of both the main Flash Attention pipeline and the split-k reduction pipeline (pipeline_flash_attn_split_k_reduce).

Additional information

• Bypasses the 4GB maxStorageBufferRange limit by leveraging the existing 64-bit indexing pipeline variants already compiled in the Vulkan backend.
• Successfully verified with a prompt containing 80,000 tokens on dual AMD Radeon RX 6800 GPUs.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - Used an AI coding assistant (Antigravity) in an assistive capacity to help identify the missing 64-bit pipeline selections in the Flash Attention code path and draft the initial patch. All changes have been manually reviewed, compiled, and tested on physical hardware.