Remove redundant attributes specifying the number of threads per warp

[FMarno]

In #1270, the SPIRV Target Env attribute was added to communicate the subgroup-size/threads per warp to triton_gen.sub_group_reduce. This now exists in parallel with triton_gpu.threads-per-warp, and also probably overrides the need for triton_gen.intel_reqd_sub_group_size.
This redundancy should be investigate properly and a removed if possible.

[FMarno]

I'm working on a POC but it'll be a while before it can land since it requires changes to upstream GPU dialect.
Triton POC: https://github.com/intel/intel-xpu-backend-for-triton/tree/finlay/poc_gpu_subgroup_size
Upstream changes: https://github.com/FMarno/llvm-project/tree/gpu_named_attrs

[FMarno]

I'm working on a POC but it'll be a while before it can land since it requires changes to upstream GPU dialect. Triton POC: https://github.com/intel/intel-xpu-backend-for-triton/tree/finlay/poc_gpu_subgroup_size Upstream changes: https://github.com/FMarno/llvm-project/tree/gpu_named_attrs

This didn't work out so well, but I'm working on a new POC now

[FMarno]

POC in progress here: https://github.com/intel/intel-xpu-backend-for-triton/tree/finlay/poc_spirv_barrier
and FMarno/llvm-project#4