sycl: Add more debug prints

[Rbiessy]

This adds more debugging prints related to the operations being run. scope_op_debug_print is introduced to print the "call ..." and "call ... done" logs in a single line.
This adds more information for all the operations and its destination and input tensors: the tensor type, its number of elements, stride sizes and whether the tensor is strided (i.e. non-contiguous) or permuted.
For mul_mat this also add some debug prints when conversations kernels are called from and to fp16, fp32 and quantized types.
This also adds debugging prints for a few operations that were missing it.

The output for llama-2-7b.Q4_0 PP and TG is added here as an example:

• llama-2-7b.Q4_0__log_bench_pp_512_no_warmup.txt
• llama-2-7b.Q4_0__log_bench_tg_1_no_warmup.txt

[qnixsynapse]

Nice.
Is it possible to extend this to backend interface functions as well such as init, set, memset, clear, reset, etc?

[Rbiessy]

Nice. Is it possible to extend this to backend interface functions as well such as init, set, memset, clear, reset, etc?

I added some more logs and updated the example in the PR description. Note I am not reusing scope_op_debug_print everywhere as I wanted to make the difference between logs coming from the llama backend (like the functions you mentioned) and logs coming from the model (like operations that are run). I added some [OP] tag in the log to distinct between the 2.

[Alcpz]

LGTM. If logs get too noisy we can implement a couple of levels to the debug variable in the future to either print ops or API calls.

[qnixsynapse]

LGTM

[Rbiessy]

I switched from using std::string to std::string_view to avoid any cost associated with std::string. As I measured the PR wasn't enough to impact performance but a small reproducer showed more instructions generated. With std::string_view that is not the case anymore, see f1 (using macro) and f2 (using scope_dbg_print) are identical: https://gcc.godbolt.org/z/5T84n5rzs
I'll wait until Monday to merge this if there are no other comments.