test: fix BFloat16 math test on the Aurora LTS stack

The gpuarrays/math/intrinsics BFloat16 subtest failed two ways:

1. Host-side: Core.BFloat16 exists on Julia 1.12 but has no Real->bf16
   constructors (those live in BFloat16s.jl), so ET(-1) threw a MethodError
   before any GPU work. Add BFloat16s as a test dep and load it in the
   worker so the CPU reference path works.

2. Codegen: _device_supports_bfloat16() is a hardware check, but the Aurora
   LTS SPIR-V stack (Khronos translator + NEO/IGC) cannot codegen native
   bfloat in generic kernels -- clamp! fails with InvalidIRError, and
   declaring SPV_KHR_bfloat16 crashes the LTS runtime. Gate the bf16 eltypes
   push on !oneL0.LTS[] so BFloat16 is only exercised as a generic element
   type off the LTS stack (storage/conversions/oneMKL bf16 are unaffected).

Author: michel2323

test/Project.toml

@@ -1,6 +1,7 @@
 [deps]
 AbstractFFTs = "621f4979-c628-5d54-868e-fcf4e3e8185c"
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
+BFloat16s = "ab4f0b2a-ad5b-11e8-123f-65d77653426b"
 Dates = "ade2ca70-3891-5945-98fb-dc099432e06a"
 Distributed = "8ba89e20-285c-5b6f-9357-94700520ee1b"
 FFTW = "7a1cc6ca-52ef-59f5-83cd-3a7055c09341"

test/runtests.jl

@@ -64,6 +64,9 @@ end
 init_worker_code = quote
     $device_claim_code
     using oneAPI, Adapt
+    # BFloat16s.jl supplies the host-side numeric conversions (e.g. BFloat16(::Real))
+    # that Core.BFloat16 lacks on its own; the testsuite's CPU reference path needs them.
+    using BFloat16s
 
     import GPUArrays
     include($gpuarrays_testsuite)
@@ -83,7 +86,12 @@ init_worker_code = quote
         append!(eltypes, [Float64, ComplexF64])
     end
     @static if isdefined(Core, :BFloat16)
-        const bfloat16_supported = oneAPI._device_supports_bfloat16()
+        # _device_supports_bfloat16() is a hardware check (bf16 conversions ext / PVC device),
+        # but the Aurora LTS SPIR-V stack (Khronos translator + NEO/IGC) cannot codegen native
+        # `bfloat` in generic kernels: any kernel that keeps a bfloat value (e.g. clamp!) fails
+        # with InvalidIRError, and declaring SPV_KHR_bfloat16 crashes the LTS runtime. So only
+        # exercise BFloat16 as a generic element type off the LTS stack.
+        const bfloat16_supported = oneAPI._device_supports_bfloat16() && !oneL0.LTS[]
         if bfloat16_supported
             push!(eltypes, Core.BFloat16)
         end