[auphelia]

Extends the hardware (board) CI: the AUP-ZU3 is now a real test board, and the MLO and external-weights builds are tested on hardware, not only built. Adds filters to run part of the HW pipeline for debugging.

Changes

Boards

• AUP-ZU3 (finn-aup-zu3) replaces the Pynq-Z1 placeholder: real agent label, credentials and reboot before testing.

New HW test types

• mlo_build (MLO): the four AUP-ZU3 variants of test_finnloop_end2end_mlo_ddr.
• ext_weights_build (Ext weights): test_end2end_ext_weights_build, cnv and tfc, now built for the AUP-ZU3.
• New helper finn.util.test.stage_hw_deployment() lets any build test hand its bitfile, plus a reference input/output, to the HW pipeline. Documented in ci/README.md.
• On the board each model runs an execute test (output compared bit-exactly with the reference) and a throughput test. The board runner now picks up whichever .bit file a package ships, not only resizer.bit.

Pipeline

• A shard now hands off the bitfiles of its passing tests even when other tests in the shard fail.
• New Jenkinsfile_HW parameters: TEST_TYPE_FILTER, MODEL_EXCLUDE, TEST_SKIP.

Fixes found by the new board tests

• make_driver.py: the external-weights repeat count was taken from a FIFO on the weight stream instead of the MVAU behind it. Conv layers got the weights once instead of once per output pixel, and the accelerator hung.
• cnv-w2a2-extw.json / specialize_layers_config_cnv.json: these used node names from before max-pool was lowered to depthwise SWG + Pool_hls. Entries landed on the wrong nodes, mismatching depthwise SWG SIMD and Pool PE, and the folded network computed wrong results. Now realigned.
• test_end2end_ext_weights_build now verifies the folded network against the source model (folded cppsim), so a broken folding config fails the build test instead of the board.