Add scratchpad run-time parameter functionality

[andrej]

Builds on the XRT/firmware scratchpad feature (#3056) to enable more efficient run-time parameters. The XRT scratchpad supports up to 32 parameters in a "StateTable" that can be written to directly from DDR on the host, copied over by the firmware into a local memory, and from there used to write to registers in the NPU partition.

The primary advantage/difference over the existing run-time parameter implementation is that these parameters are truly read from DDR. They are not hard-coded as write32s in the instruction sequence, and therefore do not require instruction sequence patching or different instruction sequences for different parameter values.

This PR abstracts the XRT scratchpad mechanism to make it more user-friendly through an aiex.parameter declaration (at the module level since parameters can be shared across devices), a aiex.sync_parameters_from_host operation that performs the copy from DDR -> command processor memory -> NPU registers/buffers, and a aiex.read_parameter operation that allows reading the parameters from the core. For the host side, a convenience class ScratchpadParameter provides an abstraction to write the parameters to the correct offset in DDR (no manual tracking of offsets, instead we can use named parameters as declared in the MLIR using aiex.parameter).

The mapping of parameter name to DDR offset/StateTable index (where the host C++ writes to and runtime sequence reads from) is exported as additional metadata to be consumed by the runtime libs in a params.txt file. The mapping of parameter name to buffers on NPU cores is kept inline in the MLIR.

    test.cpp (host code) 
          |
          |    ScratchpadParameter abstraction in test_utils
          |    ScratchpadParameter.write("foo", value)
          |
 parameter in DDR 
          |
          |    aiex.sync_parameters_from_host in runtime sequence
          |
  parameter in scratchpad
      /  |  \
    /    |    \      parameters written to NPU cores, one buffer per parameter per core
  /      |      \     (this is also done by aiex.sync_parameters_from_host)
/        |        \
                 aiex.read_parameter("foo")

C++ usage example: NPU-side (MLIR), host-side (C++)

Python usage example: NPU-side (IRON Python), host-side (Python)

There is also another mode to use this where rather than use it as run-time parameters on NPU cores, you can use it to patch DMA BD addresses, C++ example, Python example.

Note: Depends on XRT PR #9813 / XRT PR #9814 for the Python bindings