XLA:TPU `mhlo.acosh` / `mhlo.acos` cannot be translated to XLA HLO

[avik-pal]

Input MLIR:

module @reactant_acosh attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
  func.func @main(%arg0: tensor<3xf32>) -> tensor<3xf32> attributes {enzymexla.memory_effects = []} {
    %0 = chlo.acosh %arg0 : tensor<3xf32> -> tensor<3xf32>
    return %0 : tensor<3xf32>
  }
}

(This is not an issue on CPU or CUDA)

This started showing up on our CI with a recent XLA version bump https://github.com/EnzymeAD/Reactant.jl/actions/runs/17976254980/job/51130447715?pr=1690#step:20:1258.

  UNKNOWN: <unknown>:0: error: 'mhlo.acosh' op can't be translated to XLA HLO
  <unknown>:0: note: see current operation: %0 = "mhlo.acosh"(%arg0) : (tensor<2xf64>) -> tensor<2xf64>

[akuegel]

Likely related to tensorflow/tensorflow#100184
Acosh is now a real HLO op, likely you are missing a lowering for that which you now have to implement. Before, chlo.acosh was expanded to a pattern of other HLO ops.