Merge upstream→gfx11

[eble-amd]

Among the changes to torch bindings, there were some diff hunks where both upstream and local added lines. I kept both.

Finding what to merge was tricky this time. My initial attempt to stop at the first conflicting commit hit errors related to torch version and bindings. For more information, see the series of changes vllm-project#44334, vllm-project#44648, vllm-project#47128.

[eble-amd]

Manual sanity test on gfx1151:

Arguments: --model Qwen/Qwen2.5-0.5B-Instruct-AWQ --num-prompts 10 --max-model-len 4096 --kill-e
xisting-vllm-processes --input-len 3968 --output-len 128 --target-gpu-memory-gb 10 --max-num-seq
s 1 -e TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 -e FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE -e TO
RCH_BLAS_PREFER_HIPBLASLT=1
Prefill 23338.17 tokens/s; TTFT 170 ms
Decode 312.9 tokens/s (TPOT 3.20 ms)
End-to-end latency 575 ms (median)
Average CPU utilization 7%

[eble-amd]

Performance regression tests on gfx1151:

python -m pytest -v tests/kernels/quantization/test_hybrid_w4a16_perf.py

...
========================= 128 passed, 17 warnings in 436.47s (0:07:16) =========================
...

[mgehre-amd]

Thanks!