metal: Implement ROLL op

[kushagharahi]

Overview

Per #21941 the roll op was not implemented for metal backend. Falling back to CPU.

I regenerated the ops docs (using M1) as well with the following:

./build/bin/test-backend-ops support -b MTL0 --output csv > ./docs/ops/Metal.csv

python3 ./scripts/create_ops_docs.py

Testing

Post implementation llama-server does not print the the CLIP graph uses unsupported operators by the backend logs for ROLL

./build/bin/test-backend-ops -o ROLL -b MTL0

ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices
ggml_metal_library_init: using embedded metal library
ggml_metal_library_init: loaded in 0.016 sec
ggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)
ggml_metal_device_init: GPU name:   MTL0
ggml_metal_device_init: GPU family: MTLGPUFamilyApple7  (1007)
ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_device_init: GPU family: MTLGPUFamilyMetal3  (5001)
ggml_metal_device_init: simdgroup reduction   = true
ggml_metal_device_init: simdgroup matrix mul. = true
ggml_metal_device_init: has unified memory    = true
ggml_metal_device_init: has bfloat            = true
ggml_metal_device_init: has tensor            = false
ggml_metal_device_init: use residency sets    = true
ggml_metal_device_init: use shared buffers    = true
ggml_metal_device_init: recommendedMaxWorkingSetSize  =  5726.63 MB
Testing 3 devices

ggml_metal_init: allocating
ggml_metal_init: found device: Apple M1
ggml_metal_init: picking default device: Apple M1
ggml_metal_init: use fusion         = true
ggml_metal_init: use concurrency    = true
ggml_metal_init: use graph optimize = true
Backend 1/3: MTL0
  Device description: Apple M1
  Device memory: 5461 MB (5460 MB free)

ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_roll_f32', name = 'kernel_roll_f32'
ggml_metal_library_compile_pipeline: loaded kernel_roll_f32                               0x12ee0b290 | th_max = 1024 | th_width =   32
  ROLL(shift0=3,shift1=-2,shift3=1,shift4=-1): OK
  1/1 tests passed
  Backend MTL0: OK
ggml_metal_free: deallocating
Backend 2/3: BLAS
  Skipping
Backend 3/3: CPU
  Skipping
3/3 backends passed
OK

Requirements

• I have read and agree with the contributing guidelines Yes
• AI usage disclosure: Opus was used for assistance in the implementation. Verification, tests and final changes done by me.

ref #14909 too

[ngxson]

CC @ggerganov if you have a bit of time, this can significantly improve the speed of conformer-based audio models (LFM and gemma 4)