Support Float8 types

[AntonOresten]

Uses DLFP8Types as it has the exact types needed.

It could be nice to in the future support multiple narrow precision type implementations (see Reactant/DLFP8TypesExt.jl), like Float8s.jl and Microfloats.jl (which currently has a MX_E4M3 equivalent to Float8_E4M3FN), but there's no immediate need since Float8s.jl doesn't have Float8_5, nor a type for finite-only, single NaN encoding.

julia> @be CUDA.@sync run(data8; tm=128, tn=128, tk=128)
Benchmark: 380 samples with 1 evaluation
 min    247.029 μs (442 allocs: 10.141 KiB)
 median 252.549 μs (442 allocs: 10.141 KiB)
 mean   252.715 μs (442 allocs: 10.141 KiB)
 max    378.948 μs (442 allocs: 10.141 KiB)

julia> run_others(data16, nruns=5, warmup=1)
Dict{String, Vector{Float64}} with 1 entry:
  "cuBLAS" => [0.42032, 0.408192, 0.409696, 0.409504, 0.411424]

I can resolve any potential merge conflicts caused by #34.