ggml: add env var GGML_OP_OFFLOAD_MIN_BATCH

[DocShotgun]

#18530 / #17026

Makes the min_batch_size for triggering op offload configurable via env var, defaulting to the prior hardcoded value of 32 if not specified to keep current behavior intact.

This is helpful when running large MoEs with a significant size of weights stored in host buffers on CPU, causing a bottleneck when op offloading with small batches that are still larger than the default 32. The optimal value, or "break even point" here depends on characteristics of the hardware + model, and is best determined empirically (ref: #17026 (comment)).

Make sure to read the contributing guidelines before submitting a PR

[ggerganov]

The env var should be read only once upon device initiali§ation and then queried from the device context.

[DocShotgun]

Took a crack at this, let me know if you'd recommend doing anything differently.

AI-assisted with searching for the relevant code in the Metal backend and with debugging compile failures.

• For CUDA, CANN, SYCL, and Vulkan, added op_offload_min_batch_size to the device context struct. We read the env var once prior to the loop that creates the device context(s), and then assign this value to the context for each device.
• For Metal we instead add the field to the device props, which we can then fetch from the offload op check.
• dev is no longer flagged as unused in the backend offload op checks. In Metal, op was also previously flagged as unused.
• CANN had an issue where ggml_backend_cann_offload_op is declared before ggml_backend_cann_device_context. This didn't cause any problems before when device context was unused. I moved it down to roughly match the other backends.

I tested CUDA locally on a Qwen3-Coder-30B-A3B-Instruct-IQ4_XS.gguf on my 7950X + 4090 Windows machine with -b 4096 -ub 4096 and --cpu-moe and it seems to work as expected:

• GGML_OP_OFFLOAD_MIN_BATCH=50000 with 1532 tokens prompt, op offload not triggered -> PP 146.81 T/s
• GGML_OP_OFFLOAD_MIN_BATCH=64 with 1532 tokens prompt, op offload triggered -> PP 2320.88 T/s
• No env var set with 1532 tokens prompt, defaults to 32, op offload triggered -> PP 2322.76 T/s

[NeoZhangJianyu]

Looks good to me for SYCL backend part.

[0cc4m]

The Vulkan changes are fine.

[am17an]

@ggerganov merge?