Default to hardware floating-point atomics.

[pxl-th]

Default to 'unsafe' hardware floating-point atomics.
TL;DR instead of emulating them via CAS loop use hardware RWM instruction that is significantly faster.
More details: link.

E.g. assembly atomic instruction before & after this PR for the following kernel:

function ker!(x)
    @inline @atomic x[1] += 1f0
    return
end

• Before: global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] glc
• After: global_atomic_add_f32 v0, v1, s[0:1]

I'm inclined to make this a default because of huge performance increase.
On Nerf.jl benchmark this gives ~2x performance improvement and on yet-unreleased GaussianSplatting.jl 17x boost in performance matching CUDA.

However, on a per-kernel basis this can be disabled with:

@roc unsafe_fp_atomics=false f(...)

CC @luraess @OsKnoth

[luraess]

cc: @albert-de-montserrat