Metal passing all tests workload

[hikettei]

• Refactor in buffer.lisp (add allocated_size param to the buffer)
  • Use buffer-storage-size to transfer arrays #411
  • change-facet for metal and all external backends without additional code modification
• Refactor in initializer.lisp (dispatch _rng_counter for each device) Refactor: each device have independent rng_counter #410
• Fix: argmax-failing-case-repro. (one sized loop for should not get GLOBAL attribute)
• Enhancement: Global Runtime #412
• fix NATIVE=1