feat: Ops.throw

[avik-pal]

Still need to register the FFI on the cpp end. Is that the best way to do this or do we have a way of registering function calls on julia end?

julia> @code_hlo error("dimension mismatch")
AssertionError("Could not find registered platform with name: \"cuda\". Available platform names are: ")
module @reactant_error attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
  func.func @main() {
    %c = stablehlo.constant dense<true> : tensor<i1>
    stablehlo.custom_call @throw(%c) {api_version = 4 : i32, backend_config = {error_message = "dimension mismatch"}, has_side_effect = true} : (tensor<i1>) -> ()
    return
  }
}

[avik-pal]

though I am not super sure if we should do this by default. There are a lot of usecases for having the runtime error functionality separately but with this all errors will be embedded into mlir by default

[avik-pal]

julia> fn(x) = begin
           Reactant.Ops.throw("dimension mismatch")
           return x .+ 1
       end
fn (generic function with 1 method)

julia> @jit fn(Reactant.to_rarray(rand(2)))
ERROR: INTERNAL: dimension mismatch

Stacktrace:
 [1] reactant_err(msg::Cstring)
   @ Reactant.XLA ~/reactant/Reactant.jl/src/xla/Utils.jl:12
 [2] macro expansion
   @ ~/reactant/Reactant.jl/src/xla/PJRT/LoadedExecutable.jl:175 [inlined]
 [3] execute_sharded
   @ ~/reactant/Reactant.jl/src/xla/PJRT/LoadedExecutable.jl:144 [inlined]
 [4] macro expansion
   @ ~/reactant/Reactant.jl/src/Compiler.jl:1435 [inlined]
 [5] (::Reactant.Compiler.Thunk{typeof(fn), Symbol("##fn_reactant#280"), Tuple{ConcretePJRTArray{…}}, false})(args::ConcretePJRTArray{Float64, 1, 1, Reactant.Sharding.ShardInfo{Reactant.Sharding.NoSharding, Nothing}})
   @ Reactant.Compiler ~/reactant/Reactant.jl/src/Compiler.jl:1690
 [6] top-level scope
   @ ~/reactant/Reactant.jl/src/Compiler.jl:1036
Some type information was truncated. Use `show(err)` to see complete types.

[github-actions]

[JuliaFormatter] _{reported by reviewdog 🐶}

Suggested change

	location=mlir_stacktrace("throw", @__FILE__, @__LINE__)
	location=mlir_stacktrace("throw", @__FILE__, @__LINE__),

[avik-pal]

need to fix cuda before merging this