[Issue]: Improve the Iris allocator

Problem Description

Iris currently implements a very simple bump allocator with a noop deallocate. See:
https://github.com/ROCm/iris/blob/f7876ba15db4992e90184861f538526087c3288a/iris/iris.py#L266-L267

This causes out-of-memory errors and memory usage going up all the time. There were two solutions in mind (please suggest others as well):

1. Implement a better allocator in Python (from a simple free-list to more complex ones),
2. Use a Python-bindings around a C/C++ allocator (e.g., dlmalloc.cpp)

A simple free-list of tensors will likely be sufficient and is probably the easiest approach for now. Since it's still unclear how Iris will integrate with inference frameworks, the level of effort we should invest in this is uncertain. Use your best judgment when deciding how far to take the implementation. Feel free to discuss here or offline.

See initial freelist implementation here: https://gist.github.com/maawad/763c981d3caeeca012858ee45fa1803b A solution using an ordered map would be better.

We will need a C++ implementation.

Operating System

Any

CPU

Any

GPU

Any

ROCm Version

Any

ROCm Component

No response

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

The current allocator is referenced at iris/iris.py#L266-L267; start there, then compare the linked freelist gist and rocSHMEM's dlmalloc.cpp. Define the C++ allocator scope and validate that deallocation prevents the reported memory growth and out-of-memory behavior; the issue does not specify tests.