druidExitPoint does not work for meta-compilation of bytecodes

[PalumboN]

I found that the definition for pushNewArrayBytecode generated by Druid does not manage a border case well.

This bytecode need to instantiate an array, so there is a check using the objectMemory getScavengeThreshold to check if the new object fits in the new space:
https://github.com/pharo-project/pharo-vm/blob/eaa8ca205aa46fe79bd87dae2cb75244f67e743e/smalltalksrc/VMMaker/DruidJIT.class.st#L12691-L12695

The problem is that, after the comparison, there is a jump to the end of the bytecode in case of "not enough space".

This happen because the check in the interpreter uses druidExitPoint instrinsic.

	freeStart + numBytes > scavengeThreshold ifTrue: [
		self druidExitPoint.
		needGCFlag ifFalse: [self scheduleScavenge].
		self validateRoomInEdenOrDie: numBytes ].

But that intrinsic is implemented to work only with primitives:

druid/Druid/DRMetaCompilerIRGenerator.class.st

Lines 62 to 78 in 742b6c9

	DRMetaCompilerIRGenerator >> interpretDruidExitPointWith: aRBMethodNode [
	| failure primitiveReturn returnState |
	failure := self currentBasicBlock copy: 1.
	self currentPrimitiveFailureCode: failure.
	returnState := executionState copy.
	primitiveReturn := returnState baseFrame.
	primitiveReturn push: DRNullValue new. "Return something"
	"Record that this execution stack did a method return
	We will merge this one with other returned states when this method finishes"
	primitiveReturn addDeferredReturnState: {
	returnState.
	(self addFrameReturn: primitiveReturn) }
	]

That is why it is jumping at the end of the definition (for primitives, it is the way to "fail" and execute the fallback).

We need to merge druidExitPoint (for primitives) with druidForceIntepretation (for bytecodes).