Re-enable semi-concrete evaluation

[maleadt]

JuliaLang/julia#47994 fixed most cases where semi-concrete interpretation ended up with a wider result than constant propagation. It may be okay to turn on semi-concrete interpretation in GPUCompiler after 1.10.0-DEV.203. However, there are still a few known remaining cases, so you may want to keep the overload unless you are willing to optimize compilation performance at the risk of inference regression.

Originally posted by @aviatesk in JuliaLang/julia#47349 (comment)

[maleadt]

Reverted in #760, x-ref #761.

[maleadt]

Reproducer:

using Metal, Test; let # sincos
           T = Float32
           N = 4
           arr = rand(T, N)
           bufferA = MtlArray(arr)
           bufferB = MtlArray(arr)
           function intr_test3(arr_sin, arr_cos)
               idx = thread_position_in_grid().x
               sinres, cosres = sincos(arr_cos[idx])
               arr_sin[idx] = sinres
               arr_cos[idx] = cosres
               return nothing
           end

           Metal.@sync @metal threads = N intr_test3(bufferA, bufferB)
           @test Array(bufferA) ≈ sin.(arr)
           @test Array(bufferB) ≈ cos.(arr)
       end

[christiangnrd]

Strangely the reproducer does not work on the buildkite runners, but has reproduced locally on every machine I've tried (M1 like the runners, M2, M2 Max)