fix: stabilize tassemble f32->f16 path and add A3 board validation

[TaoTao-real]

Summary

• Relax pto.tassemble verifier to allow src=f32 -> dst=f16/bf16 for A3 TINSERT cast-quant path.
• Rework test/samples/Assemble/assemble.py to generate an A3-legal TINSERT flow and keep emitted-pattern regression.
• Refresh test/samples/Assemble/board_validation harness/files and fix local address overlap in generated kernel test artifact.
• Extend board run script behavior (golden mode/device selection) for stable board execution.

Motivation

• tassemble lowering targets TINSERT(dst_mat, src_acc, ...) on A3.
• PTO-ISA constraints require ACC(f32) -> MAT(f16/bf16) cast-quant form in this path.
• Existing strict same-dtype verifier and board harness blocked real board validation.

Design

• TAssembleOp::verify() now accepts either:
  • same element type, or
  • src=f32 with dst=f16/bf16.
• Assemble sample builds a valid pipeline around tassemble and checks emitted C++ patterns.
• Board validation case is aligned with current generated kernel signatures and compares golden_v4.bin vs v4.bin.

Testing

• Local targeted:
  • test/samples/runop.sh -t Assemble (pass)
• Local full sample regression:
  • test/samples/runop.sh all (OK=169 FAIL=0 SKIP=1)
• A3 board:
  • GOLDEN_MODE=npu ACL_DEVICE_ID=2 .../run_a3_board_case.sh test/samples/Assemble/board_validation
  • Passed run artifacts:
    • /home/zhongxuan/ptoas_board_runs/board_validation_20260314_215720

Risk / Rollback

• Risk is limited to tassemble dtype verification and the new Assemble sample/board harness.
• Rollback by reverting this PR; no global pass defaults were changed.