Fix out-of-bounds write in point query instance stack

[stefanatwork]

The point query variant of instance_id_stack::push() guarded the instance stack capacity only with an assert, which is compiled out in release builds. It then unconditionally wrote the instance ID and both 4x4 transform matrices at context->instStackSize and always returned true.

Since rtcPointQuery() recurses into nested instances and all call sites gate recursion on the return value of push(), the missing runtime check meant a scene nested deeper than RTC_MAX_INSTANCE_LEVEL_COUNT drove instStackSize past the fixed-size arrays of the caller provided RTCPointQueryContext, writing instance transform data out of bounds. With the default RTC_MAX_INSTANCE_LEVEL_COUNT of 1, an instance of an instance is enough to corrupt memory past the context; deeper nesting scales the write distance arbitrarily.

Add the same runtime capacity check that the intersect/occluded variant of push() already performs: return false before any write when the stack is full. This makes point queries silently ignore instances nested deeper than RTC_MAX_INSTANCE_LEVEL_COUNT, which is the behaviour already documented for instancing and already implemented for ray queries. The existing call sites handle a false return by skipping the instance without a matching pop(), so the stack stays balanced.