[luca24balboni]

Summary

This PR brings the magia-v4 cluster programming model into magia_isle. The PULP cluster gets its own Event Unit and dispatches work through it, and the control core gets a direct path to L1.

Private cluster Event Unit:
The PULP cluster has its own Event Unit, with barriers, HW mutex, SW events and a dispatch FIFO for fork/join.
It gates each cluster core's clock: a core sleeps on cv.elw and wakes on an event.
Cluster cores reach it over a direct link; the control core reaches it through a memory-mapped window on the OBI crossbar.

Simpler dispatch:
The cluster control registers are reduced to fetch enable, boot address, task/data, start, done and return value.
Start and done go through the cluster Event Unit.

Direct L1 link for the control core:
A generic core_data_demux sends control-core accesses to L1 over a dedicated HCI port, to the EU over the direct link, and everything else to the OBI crossbar.
The OBI crossbar's L1 port now serves remote accesses and the private stack and reserved region.
EU accesses of the control core are serialized like cv.elw, so a pipelined load cannot overtake a blocking EU wait; it is a demux parameter, off for the cluster cores.

Address map:
New windows for the cluster EU after the collective register
The private reserved region now starts after cluster_EU.

Dependencies:
cv32e40x is bumped to lb/operand_forwarding, which forwards the third XIF operand like the first two.

Tests and SW
The PULP runtime (crt0, cluster_utils, pmsis.h) and the existing cluster tests (hello_pulp, hello_redmule_pulp, hello_spatz_pulp, fpu_cluster_test, hetero_mesh_test) are updated to the new model.

New tests:
parallel_groups and vector_add (tile), hello_pulp_barrier (mesh), on Questa and Verilator.

Misc
Architecture image updated to MAGIA v4 (control demux, collectives, cluster EU); the old image is removed.

[yvantor]

I think here we should mention #82 since the topic looks very similar (and also the name of the issue should be updated accordingly to V4 as far as I understand)

[yvantor]

I think here we should mention #82 since the topic looks very similar (and also the name of the issue should be updated accordingly to V4 as far as I understand)

I confirm this addresses #82