loader: increase async upload staging buffer to 4 MiB

[cl0ckt0wer]

Overview

On Windows, the load uses alignment == 1, so the async upload staging buffer is currently 1 MiB.

Additional information

Benchmarks are WARM READS.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - Codex assisted with code exploration, benchmarking, and PR preparation. I manually reviewed the change and benchmark results.

Benchmark

Environment:

• Model: qwen3.6:35b GGUF blob
• Loader payload: 21.45 GiB
• GPU/backend: AMD Radeon RX 7900 XTX, Vulkan
• Command shape: llama-cli --no-mmap -ngl 999
• Runs: warm-cache load timing, first 1 MiB warmup excluded

Chunk	Runs	Avg load	Median	Avg throughput
1 MiB	5	4960.94 ms	4928.89 ms	4.32 GiB/s
2 MiB	5	3687.27 ms	3735.25 ms	5.82 GiB/s
4 MiB	5	3148.23 ms	3102.45 ms	6.81 GiB/s
8 MiB	5	3246.70 ms	3254.84 ms	6.61 GiB/s
16 MiB	5	3370.00 ms	3293.10 ms	6.37 GiB/s
32 MiB	2	3264.91 ms	3255.14 ms	6.57 GiB/s
64 MiB	2	3286.43 ms	3203.69 ms	6.53 GiB/s

Validation

• git diff --check
• Windows Vulkan build:
  • cmake -S . -B C:\tmp\llama-pr-build -DGGML_VULKAN=ON -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_SERVER=OFF
  • cmake --build C:\tmp\llama-pr-build --config Release --parallel 8 --target llama-bench