SYCL: Refactor ggml_sycl_compute_forward

[qnixsynapse]

I originally thought I wouldn't do this, but the original code is quite restrictive imo. It assumes that all operations need only src0, src1 and dst tensors, which isn't the case for every OP(also, why are we passing them separately?). This change will allow easy addition of tensor operations to the backend as the dst tensor itself contains everything needed to perform the operation. I haven't changed the kernels except for a few function signature changes, as that's outside the scope of this refactoring. Those can be refactored later if needed.

[Alcpz]

Code looks simpler now. Thanks for the change.

Edit: Will wait for other to give a review.

[qnixsynapse]

@Alcpz Thanks for your review.
Tagging @Rbiessy @NeoZhangJianyu @airMeng manually to get their thoughts on this.