metal: support gated delta recurrent snapshots for MTP rollback

[i386]

Overview

Adds Metal backend support for the GGML_OP_GATED_DELTA_NET recurrent snapshot layout used by the MTP rollback path in ggml-org#22673.

The MTP path widens the recurrent state input to (S_v * S_v * H, K, n_seqs), where K = 1 + n_rs_seq. Slot 0 remains the current input state. For K > 1, the Metal kernel now writes the last min(n_tokens, K) per-token recurrent states into the trailing snapshot slots using the same slot mapping as the CPU/CUDA/Vulkan implementations:

target_slot = t - (n_tokens - K)

For K == 1, behavior remains the existing final-state-only output.

This also moves the snapshot slot count into the Metal kernel args instead of specializing the pipeline by K, so the same gated-delta pipeline can handle different recurrent snapshot counts.

Additional information

This is intended to complete the Metal side of the hybrid Qwen recurrent rollback support in ggml-org#22673.

Validated locally on Apple Metal:

cmake -S /tmp/llama-mtp-metal -B /tmp/llama-mtp-metal/build-metal \
  -DGGML_METAL=ON \
  -DLLAMA_BUILD_TESTS=ON \
  -DLLAMA_BUILD_EXAMPLES=ON \
  -DCMAKE_BUILD_TYPE=Release

cmake --build /tmp/llama-mtp-metal/build-metal --config Release --parallel 8 \
  --target llama-server test-backend-ops

/tmp/llama-mtp-metal/build-metal/bin/test-backend-ops test -b MTL0 -o GATED_DELTA_NET

Result:

GATED_DELTA_NET: 30/30 tests passed
Backend MTL0: OK

I also ran end-to-end llama-server MTP benchmarks on an Apple M1 Ultra using unsloth/Qwen3.6-27B-MTP-GGUF Q8_0, Metal backend, 9 prompt corpus, 192 generated tokens per prompt, prompt cache disabled, and one fresh server state per condition.

Mode	Runs	Mean wall s	Mean tok/s	Mean speedup	Accept
baseline	3	105.48	16.38	1.000x	n/a
mtp_w2	3	78.87	21.91	1.337x	98.50%
mtp_w3	3	74.11	23.32	1.423x	97.37%

No outliers over 5% from each mode's median were observed in that fresh-state run.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - AI assistance was used to help prepare and validate this patch and to draft this PR text. The submitted changes and PR description have been reviewed and edited by the human contributor before submission.