MAGIA SDK v3 aligned with RTL

[FrancescoConti]

Removed compiler flag that broke everything -wl,--gc-sections (no problem in bare metal, destroys in SDK).

• CV32E40P with toolchain OpenHW both in Ctrl Core and in Cluster (Riscy not tested)
• Spatz with Spatz LLVM (module on mamuthones?)

PULP cluster -> printf in cluster is a clusterfuck (not an issue!).

[FrancescoConti]

gvsoc ok.
RTL still has problems: old tests using spatz use non-inititalized .rodata that in theory should be init'ed. Tests using only PULP cluster are ok. Problem arises when cluster cores are not present. lb from non-initialized address that in RTL causes issues.

GVSOC + RTL all use CV32E40P. Toolchain of Giuseppe. As of fix of last saturday this is a release candidate: passes all GCC tests + dedicated ones for changes. No issues seen yet in linalg kernels, no perf regressions seen yet. HW loops were very suboptimal, now they are the best possible with this version of GCC (GCC 14 - more rebasable on recent GCC).

Toolchain on mamuthones: module ready, can be moved to common location. In chips-it, good to use same version.

[FrancescoConti]

CI ok, currently rebasing on main. Should be mergeable this week.
Realigning CV32E40P/X on control core.