Use host USM for scalar oneMKL results

[michel2323]

Closes #234.

The scalar-returning oneMKL BLAS wrappers hand the library a one-element device oneArray and copy the value back:

result = oneArray{$ret_type}([0])   # host vector -> device (H2D)
$fname(sycl_queue(queue), n, x, stride(x,1), result)
res = Array(result)                 # device -> host (D2H)
return res[1]

The two transfers dominate the call. Since we now support host-USM-backed arrays (oneArray{T,N,oneL0.HostBuffer} with host-side getindex/setindex!), the result can live in host USM instead: oneMKL writes straight into host-visible memory and we read it with a plain load. The C wrappers wait_and_throw() before returning, so the value is ready as soon as the ccall is.

New helper in lib/mkl/utils.jl:

function scalar_result(::Type{T}) where {T}
    res = oneArray{T, 1, oneL0.HostBuffer}(undef, 1)
    res[1] = zero(T)
    return res
end

Applied to all five sites in lib/mkl/wrappers_blas.jl: nrm2, dot/dotc/dotu, asum, iamax, iamin.

This also makes #234 moot as originally written — a dedicated ZeHostRef type isn't needed, since host-USM oneArrays cover the use case and oneL0.host_alloc handles the raw-pointer case.

Performance

nrm2 over a 4096-element Float32 vector, Arc A750:

	µs/call
device buffer + D2H copy	3079.5
host USM	331.5

~9.5x. It's pure call overhead, so the win is independent of vector length.

Testing

Arc A750, NEO 26.18.38308, oneMKL 2025.3.0, Julia 1.12.6:

• test/runtests.jl onemkl — 1112 pass, 48 broken, SUCCESS
• test/runtests.jl gpuarrays/linalg array — 2492 pass, SUCCESS
• All five wrappers spot-checked against LinearAlgebra for Float32/Float64/ComplexF32/ComplexF64

Notes

Two pre-existing things I left alone, happy to fold in if you'd rather:

• iamin is typed x::StridedArray{$elty} where every sibling uses oneStridedArray (wrappers_blas.jl:838).
• These wrappers ignore the int status the C functions return, so a oneMKL exception surfaces as a silent zero rather than an error. That's why scalar_result zero-initializes rather than leaving the buffer undef — it preserves the current behaviour on that path instead of returning garbage.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 80.33%. Comparing base (8fd0eff) to head (6badafa).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #604      +/-   ##
==========================================
+ Coverage   80.24%   80.33%   +0.09%     
==========================================
  Files          50       50              
  Lines        3487     3488       +1     
==========================================
+ Hits         2798     2802       +4     
+ Misses        689      686       -3     

☔ 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.