[losewayy]

Overview

ref #17654

Remove GGML_OP_DIAG_MASK_INF, GGML_OP_DIAG_MASK_ZERO and the four ggml_diag_mask_* functions for v1.0.0, as discussed in #17654. This PR targets the #29801 branch.

This also removes their CPU, CUDA, Vulkan, SYCL, OpenCL and CANN kernels, the backward cases, the DIAG_MASK_INF case in the Vulkan debug check, the test-backend-ops cases and the docs/ops entries. No model in the repo uses these ops.

To migrate:

• ggml_diag_mask_zero(ctx, a, 0) -> ggml_tri(ctx, a, GGML_TRI_TYPE_LOWER_DIAG), for a contiguous square F32 a, not in place
• ggml_diag_mask_inf before a softmax -> pass an additive mask to ggml_soft_max_ext, with 0 for visible and -INFINITY for masked positions

Additional information

Removing these ops shifts the IDs of later ops, as the ADD1 removal in #29801 already does. I left the RPC and backend API versions unchanged so they can be bumped once for the whole branch. Note that the RPC handshake does not check the patch version, so a patch bump alone would not catch a mismatch.

Validation

test-backend-ops test -b CPU -o DIAG,TRI,FILL,SOFT_MAX,SOFT_MAX_BACK
test-backend-ops test -o DIAG,TRI,FILL,SOFT_MAX,SOFT_MAX_BACK

• All supported cases pass on CPU, NVIDIA Vulkan and AMD Vulkan.
• The RPC multi-server test passes with the client and servers built from this branch.
• The OpenCL backend and ggml-cuda.cu compile. SYCL and CANN were not built, the changes there are removals only.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - Codex helps with implementation and code review.