[microNPU] Add a pass to move allocate nodes to the outer scope

[lhutton1]

Adds a pass called HoistAllocates to move allocate nodes to the top of the body of the main function. In doing so, it opens the door to other optimizations that need to swap the ordering of external calls.

Pass illustration:
(before)

allocate {
    extern_call(...)
    allocate {
        extern_call(...)
    }
}

(after)

allocate {
    allocate {
        extern_call(...)
        extern_call(...)
    }
}

cc @manupa-arm @NicolaLancellotti @ekalda @GeorgeARM

[manupak]

Thanks @lhutton1 ! for this backend pass.

I did an initial pass.

[lhutton1]

Thanks for the review @manupa-arm, I'll respond here since they are related :) Since this pass runs in _lower_to_tir, the pass will only ever see individual prim funcs with the "main" global function. The refactor (#10599) didn't cover this part of the compiler as it would have made it much more complex, so all the passes here: https://github.com/apache/tvm/blob/main/python/tvm/relay/backend/contrib/ethosu/tir/compiler.py#L76 are still function -> function, while the module -> module part is orchestrated by the LowerToTIR pass. We could add the functionality, although it won't be needed until we choose to refactor this part of the lowering to be module->module as well

[manupak]

Thanks @lhutton1 ! I see.

It might be worth mentioning in the docs for the pass that this pass is only meant to be run in LowerToTIR() composite pass.
We might needs some checks to that effect with tests -- i.e. checking whether the module has a single PrimFunc called "main" and if not throwing a message that the pass is only supposed to run in LowerToTIR() where the above requirement is satisfied.

[NicolaLancellotti]

LGTM!

[ekalda]

Good work @lhutton1!

[manupak]

LGTM bar @ekalda's comments.

[manupak]

LGTM!

[manupak]

Thanks! @lhutton1 @ekalda @NicolaLancellotti