SYCL: implement GGML_OP_SET + fix debug src count + align F16 assert

[GittyBurstein]

Summary

This PR adds a SYCL backend implementation for GGML_OP_SET, fixes the debug
source-count for the operator (2 sources: src0, src1), and aligns the
type-assert to allow F16 when GGML_SYCL_F16 is enabled.

Motivation

• GGML_OP_SET is frequently used as a two-step operation:
  1. base copy of src0 → dst
  2. overlay/set from src1 with offset/strides.
• The previous debug trace declared num_src=1, which is misleading.
• device_supports_op() already whitelists F16 (when enabled), so the
  implementation assert should match that contract.

What’s in this PR

• New/complete SYCL path for GGML_OP_SET:
  • Fast path for contiguous buffers (memcpy/queue copy).
  • Generic kernel for strided / non-contiguous layouts.
  • Two-phase logic (base copy, then overlay) on the same queue.
• Debug fix: scope_op_debug_print(..., /*num_src=*/2).
• Type assert alignment:
  • Default: F32 and I32.
  • With GGML_SYCL_F16: allow F16 as well (copy is size-aware, so safe).

Implementation Notes

• The code uses ggml_type_size(dst->type) to do raw-size-aware copies.
• No barriers are required for in-order queues (both phases reside on the same
  SYCL queue). If desired, a submit barrier can be added as a conservative
  measure.
• Kept behavior consistent with other element-wise ops and device_supports_op().

Tests

• Built with -DLLAMA_SYCL=ON and ran:
  • ./bin/test-backend-ops test -b CPU -o SET ✅
  • ./bin/test-backend-ops test -b SYCL0 -o SET ✅ (OpenCL GPU)
• Optional: built with GGML_SYCL_F16 to validate the F16 path.

Performance

• Contiguous paths use direct copies for best throughput.
• Strided paths use a simple parallel kernel; no observed regressions vs. CPU.

Checklist

• Compiles with and without GGML_SYCL_F16
• Tests pass on CPU and SYCL0
• Style/formatting applied (clang-format)
• No public API changes

Thanks for reviewing!