Preusser/mvu harden

[preusser]

This PR improves the usability and timing closure of the DSP-based dot-product kernels by:

• Pipelining the terminal reduction trees across parallel partial accumulations,
• Allowing to trade the available bit allocation space between weights and activations,
• Avoiding assign statements for the clock selection with difficult simulation scheduling semantics,
• Allowing very wide accumulations even spilling the high lane out of the DSP datapath, and
• Having the core DSP pipeline run freely without a high-fanout enable signal using zero-product insertion instead.