Add EXLA to enable use of at least CPU-based ML

From working with Paulo Valente and Jonatan Klosko it seems we
need to run from unreleased versions right now. Not sure what
fix it is we need but that was the case when I tried to run
Whisper.

mix.exs

@@ -11,6 +11,7 @@ defmodule NervesLivebook.MixProject do
 
   # See the BlueHeron repository for the boards that it supports.
   @ble_targets [:rpi0, :rpi3, :rpi3a]
+  @xla_targets [:rpi4, :rpi5]
 
   # Instruct the compiler to create deterministic builds to minimize
   # differences between firmware versions. This helps delta firmware update
@@ -110,6 +111,9 @@ defmodule NervesLivebook.MixProject do
       {:vega_lite, "~> 0.1"},
       {:vintage_net_wifi, "~> 0.12.5", targets: @all_targets},
       {:vintage_net_qmi, "~> 0.4.1", targets: @all_targets},
+      {:axon, "~> 0.7", targets: @xla_targets},
+      {:bumblebee, "~> 0.6", targets: @xla_targets},
+      {:exla, "~> 0.9", targets: @xla_targets},
 
       # Nerves system dependencies
       {:nerves_system_rpi, "~> 1.28", runtime: false, targets: :rpi},