Enable SPV_KHR_float_controls2 when the device accepts it

[vchuravy]

Draft until JuliaGPU/GPUCompiler.jl#1007 is released (see below). Same change as JuliaGPU/KernelAbstractions.jl#876.

Base SPIR-V can't carry LLVM's contract fast-math flag. So muladd reaches the driver as a separate OpFMul + OpFAdd, the SPIR-V back-end marks every kernel ContractionOff, and the driver never emits an FMA (JuliaGPU/KernelAbstractions.jl#865). SPV_KHR_float_controls2 carries the flag as FPFastMathMode AllowContract, and both SPIR-V back-ends emit that when the extension is enabled.

Once enabled, though, the back-ends declare the extension in every float kernel, and drivers that don't accept it reject the module. So this PR:

• adds a spirv_extensions device property (CL_DEVICE_SPIRV_EXTENSIONS_KHR, from cl_khr_spirv_queries; empty when the device doesn't support the query);
• defaults extensions in the compiler config to the OPTIONAL_SPIRV_EXTENSIONS (currently just SPV_KHR_float_controls2) that the device lists. Programs built through spirv2clc don't get it, and explicitly passed extensions replace the default;
• adds a missing comma in propertynames(::Device), which made :non_uniform_work_group_support :extensions parse as a range, so calling it threw.

Dependencies:

• Drivers: none of the devices I tested report the extension yet (PoCL 7.2, Intel CPU runtime, rusticl, AMD APP, NVIDIA), so for them this PR changes nothing. Enable SPV_KHR_float_controls2 for the CPU devices pocl/pocl#2393 enables it for PoCL's CPU devices with LLVM >= 23.
• GPUCompiler: the LLVM back-end produces invalid SPIR-V for float kernels without contract operations once the extension is on. It omits the FloatControls2 capability that its implicit FPFastMathDefault needs. SPIR-V: declare FloatControls2 for kernels with SPV_KHR_float_controls2 GPUCompiler.jl#1007 works around this. Before merging, this PR should raise the GPUCompiler compat to the release that contains it.

Testing:

• spirv_extensions is added to the device-info test, and a new "fast-math contraction" test checks three things: AllowContract is emitted exactly when the config enables the extension; the extension is enabled when the device reports it; and extensions = String[] turns it off.
• device and execution tests on every platform on this machine (PoCL, PoCL-C, Intel CPU, AMD APP, NVIDIA, rusticl): the same results as main, plus the new tests. The existing failures match main: Intel device 1, AMD device 2, AMD execution 1, rusticl execution 1.
• Full suite, with a PoCL build of Enable SPV_KHR_float_controls2 for the CPU devices pocl/pocl#2393 (LLVM 23.1.1) and SPIR-V: declare FloatControls2 for kernels with SPV_KHR_float_controls2 GPUCompiler.jl#1007: 30,515 passes versus 30,509 on main with the same PoCL, where the extra 6 are the new tests. Both runs crash in the same two files, pocl/intrinsics and poclc/intrinsics, with illegal instructions from that PoCL dev build.
• End to end on that PoCL: @opencl of y[i] = muladd(a[i], x[i], y[i]) compiles to vfmadd and gives results bit-identical to fma.(a, x, y).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.08%. Comparing base (7b8d2c4) to head (8e929c9).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #562      +/-   ##
==========================================
+ Coverage   86.06%   86.08%   +0.01%     
==========================================
  Files          17       17              
  Lines        1615     1617       +2     
==========================================
+ Hits         1390     1392       +2     
  Misses        225      225              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.