Fix pulp_zero_mem to avoid illegal HW loop

[marchioa]

pulp_zero_mem zeros a memory region with the following loop:

  for (int i=0; i<(size>>2); i++) {
    *((v4u *)pBuffer) = (v4u){0,0,0,0};
    MemoryFence();
    pBuffer+=4;
  }

This code may be translated into assembly similar to

lp.setup  0x0, a1, 0x4
p.sw  a3, 4(a0!)

On cores with more than one pipeline stage, this sequence can result in an illegal hardware loop configuration. In particular, one of the hardware loop constraints
requires the loop body to contain more than one instruction. Additionally, hardware loops do not allow memory ordering instructions (fence, fence.i) in the loop body.

The presence of MemoryFence() appears to prevent the compiler from correctly deriving the hardware loop properties, potentially leading to an invalid loop configuration. While this issue does not manifest in GVSoC simulation, it may cause unexpected behavior in RTL simulations.

A potential solution is to remove the MemoryFence() from the loop. Doing so allows the compiler to correctly generate a valid hardware loop, typically inserting a nop before the store instruction and thus avoiding the illegal configuration.

This PR removes the MemoryFence() from the main loop in pulp_zero_mem to prevent the generation of illegal hardware loops.